1107018803

9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0
Applied Optimization Methods for Wireless Networks
There is a growing trend in applying optimization approaches to study wireless

networks. This book is written to serve this need and is mainly targeted to
graduate students who are conducting theoretical research in wireless net-
works using optimization-based approaches. Specifically, the goal of this book
is to offer a course to graduate students by demonstrating a collection of
mathematical tools (with a focus on optimization techniques) and showing
how they can be used to address some challenging problems in wireless net-
works. Each chapter starts with a brief overview of a particular optimization
technique and is followed by a comprehensive coverage of a case study from
wireless networking. The choice of the case study in each chapter reflects state-
of-the-art wireless network research rather than discussing some classic but
outdated problems that are no longer of current research interest. There is a
strong emphasis on enhancing graduate students skills in mathematical mod-
eling, in particular, problem formulation, reformulation, and transformation.
For instructional purposes, each chapter contains a set of problems that are
suitable for homework exercises. A solution manual is available to instruc-
tors. PowerPoint slides for each chapter are available to both students and
instructors.
Y. Thomas Hou is a Professor in the Bradley Department of Electrical and

Computer Engineering, Virginia Tech, Blacksburg, Virginia, USA.
Yi Shi is a Research Scientist at Intelligent Automation Inc., Rockville,
Maryland, USA and an Adjunct Assistant Professor in the Bradley Department
of Electrical and Computer Engineering, Virginia Tech, Blacksburg, Virginia,
USA.
Hanif D. Sherali is a University Distinguished Professor Emeritus in the Grado
Department of Industrial and Systems Engineering, Virginia Tech, Blacksburg,
Virginia, USA. He is an elected member of the U.S. National Academy of
Engineering.
Applied Optimization
Methods for Wireless
Networks
Y. Thomas Hou
Yi Shi
Hanif D. Sherali
University Printing House, Cambridge CB2 8BS, United Kingdom
Published in the United States of America by Cambridge University Press, New York
Cambridge University Press is part of the University of Cambridge.

It furthers the Universitys mission by disseminating knowledge in the pursuit of
education, learning, and research at the highest international levels of excellence.
www.cambridge.org
Information on this title: www.cambridge.org/9781107018808
Cambridge University Press 2013
This publication is in copyright. Subject to statutory exception
and to the provisions of relevant collective licensing agreements,
no reproduction of any part may take place without the written
permission of Cambridge University Press.
First published 2013
Printed in the United Kingdom by TJ International Ltd. Padstow Cornwall
A catalog record for this publication is available from the British Library
Library of Congress Cataloguing in Publication Data
ISBN 978110701880
Cambridge University Press has no responsibility for the persistence or accuracy of
URLs for external or third-party internet websites referred to in this publication,
and does not guarantee that any content on such websites is, or will remain,
accurate or appropriate.
To our parents
and
Our wives Tingting, Meiyu, and Semeen
Contents
Preface page xi
Acknowledgments xiv
Copyright permissions xvi
1 Introduction 1
1.1 Book overview 1
1.2 Book outline 3
1.3 How to use this book 7
Part I Methods for Optimal Solutions 9

2 Linear programming and applications 11
2.1 Review of key results in linear programming 11
2.2 Case study: Lexicographic max-min rate allocation and node lifetime
problems 13
2.3 System modeling and problem formulation 15
2.4 A serial LP algorithm based on parametric analysis 20
2.5 SLP-PA for the LMM node lifetime problem 25
2.6 A mirror result 27
2.7 Numerical results 30
2.8 Chapter summary 34
2.9 Problems 36
3 Convex programming and applications 38

3.1 Review of key results in convex optimization 38
3.2 Case study: Cross-layer optimization for multi-hop MIMO networks 40
3.3 Network model 41
3.4 Dual problem decomposition 46
3.5 Solving the Lagrangian dual problem 48
3.6 Constructing a primal optimal solution 50
viii Contents

3.9 Problems 58
4 Design of polynomial-time exact algorithm 61

4.1 Problem complexity vs. solution complexity 61
4.2 Case study: Optimal cooperative relay node assignment 62
4.3 Cooperative communications: a primer 62
4.4 The relay node assignment problem 65
4.5 An optimization-based formulation 67
4.6 An exact algorithm 69
4.7 Proof of optimality 78
4.8 Numerical examples 82
4.10 Problems 89
Part II Methods for Near-optimal and Approximation Solutions 93

5 Branch-and-bound framework and application 95
5.1 Review of branch-and-bound framework 95
5.2 Case study: Power control problem for multi-hop
cognitive radio networks 100
5.3 Mathematical modeling 101
5.4 Problem formulation 108
5.5 A solution procedure 110
5.8 Problems 119
6 Reformulation-linearization technique and applications 122

6.1 An introduction of reformulation-linearization
technique (RLT) 122
6.2 Case study: Capacity maximization for multi-hop cognitive radio
networks under the physical model 125
6.3 Mathematical models 126
6.4 Reformulation 129
6.5 A solution procedure 131
6.8 Problems 146
7 Linear approximation 148

7.1 Review of linear approximation for nonlinear terms 148
7.2 Case study: Renewable sensor networks with wireless energy
transfer 151
ix Contents
7.3 Wireless energy transfer: a primer 153

7.4 Problem description 154
7.5 Renewable cycle construction 156
7.6 Optimal traveling path 162
7.7 Problem formulation and solution 164
7.8 Construction of initial transient cycle 174
7.11 Problems 179
8 Approximation algorithm and its applications - Part 1 191

8.1 Review of approximation algorithms 191
8.2 Case study: The base station placement problem 192
8.3 Network model and problem description 194
8.4 Optimal flow routing for a given base station location 196
8.5 Search space for base station location 197
8.6 Subarea division and fictitious cost points 199
8.7 Summary of algorithm and example 202
8.8 Correctness proof and complexity analysis 204
8.11 Problems 209
9 Approximation algorithm and its applications Part 2 211

9.1 Introduction 211
9.2 Case study: The mobile base station problem 212
9.3 Problem and its formulation 213
9.4 From time domain to space domain 215
9.5 A (1 )-optimal algorithm 223
9.8 Problems 241
Part III Methods for Efficient Heuristic Solutions 243

10 An efficient technique for mixed-integer optimization 245
10.1 Sequential fixing: an introduction 245
10.2 Case study: Spectrum sharing for cognitive radio networks 246
10.3 Mathematical modeling and problem formulation 247
10.4 Deriving a lower bound 253
10.5 A near-optimal algorithm based on sequential fixing 254
10.8 Problems 260
x Contents
11 Metaheuristic methods 262

11.1 Review of key results in metaheuristic methods 262
11.2 Case study: Routing for multiple description video over wireless ad hoc
networks 264
11.3 Problem description 265
11.4 A metaheuristic approach 271
11.7 Problems 280
Part IV Other Topics 281

12 Asymptotic capacity analysis 283
12.1 Review of asymptotic analysis 283
12.2 Capacity scaling laws of wireless ad hoc networks 284
12.3 Case 1: Asymptotic capacity under the protocol model 287
12.4 Case 2: Asymptotic capacity under the physical model 295
12.5 Case 3: Asymptotic capacity lower bound under the generalized
physical model 300
12.7 Problems 313
References 316
Index 327
Preface
Reasons for writing the book In recent years, there is a growing trend
in applying optimization approaches to study wireless networks. Such an
approach is usually necessary when the underlying goal is to pursue funda-
mental performance limits or theoretical results. This book is written to serve
this need and is mainly targeted to graduate students who are conducting the-
oretical research in wireless networks using optimization-based approaches.
This book will also serve as a very useful reference for researchers who wish
to explore various optimization techniques as part of their research methodolo-
gies.
To prepare a graduate student in either electrical and computer engineer-
ing (ECE) or computer science (CS) to conduct fundamental research in wire-
less networks, an ideal roadmap would include a series of graduate courses in
operations research (OR) and CS, in addition to traditional communications
and networking courses in ECE. These OR and CS courses would include
(among others) linear programming, nonlinear programming, integer program-
ming from OR, and complexity theory and algorithm design and analysis
from CS. Today, these courses are typically offered as core courses within the
respective disciplines. Instructors in OR and CS departments typically have
little knowledge of wireless networks and are unable to make a connection
between the mathematical tools and techniques in these courses and problem-
solving skills in wireless networks. ECE/CS students often find it difficult to
see how these courses would benefit their research in wireless networks. Due
to this gap between teaching scopes and learning expectations, we find that
the learning experience of our ECE/CS students in these courses is passive (or
blind) at best, as they do not have a clear picture of how these courses will
benefit their research.
An approach to bridge this gap is to offer a course that reviews a collection of
mathematical tools from OR and CS (with a focus on optimization techniques)
and shows how they can be used to address some challenging problems in
wireless networks. This book is written for this purpose.
xii Preface
Each chapter in this book starts with a brief pointer to the underlying opti-
mization technique (with references to tutorials or textbooks so that students
can do an in-depth study in a formal course or on their own). The chapter then
immediately delves into a detailed case study in wireless networks to which
the technique will be applied. The focus in each chapter is to show how the
underlying technique can be used to solve a challenging problem in wireless
networks. To achieve this goal, we offer details on how to formulate a research
problem into a formal optimization model, reformulate or transform it in order
to improve mathematical tractability, and apply the underlying optimization
technique (with necessary customizations that are specific to the underlying
problem) to derive an optimal or near-optimal solution.
We have taught this course a number of times to ECE and CS graduate stu-
dents at Virginia Tech, using chapters from this book. The response from the
students has been overwhelmingly positive. In particular, we find that:
For a graduate student (regardless of whether she has taken related OR or

CS courses), this course opens a new landscape or perspective on what opti-
mization techniques are available and how they can be applied to solve hard
problems in wireless networks;
For those graduate students who are currently taking or will take the afore-
mentioned OR and CS courses, this course will help them better appreciate
the mathematical techniques in such OR and CS courses. The student will
also have a better purpose and a stronger motivation when she takes these
core courses in her future study.
We recognize that a single-volume book cannot possibly cover all tech-

niques exhaustively. Neither is it our intention to cover everything in one book.
Nevertheless, we have organized this book into four parts, where every chapter
focuses on a single technique. We hope this organization will serve our pur-
pose of offering a first course on this important subject of Applied Optimiza-
tion Methods in Wireless Networks. Our experience shows that after taking this
course, students become substantially more mature mathematically. Most of
them are able to consciously develop their learning paths into many areas in
OR and CS not covered in this book in order to further expand their own math-
ematical capabilities. This is an important ingredient in their life-long learning
and discovery.
Finally, the idea of having a book that offers a systematic coverage of opti-
mization techniques and their applications in wireless networks is a very nat-
ural one. Unfortunately (and quite surprisingly), after a rather thorough sur-
vey of the market (when we presented our initial proposal to our publisher),
we found that there were hardly any such books available. The closest book
that we can find that by Dimitri Bertsekas: Network Optimization: Continuous
xiii Preface
and Discrete Models (Athena Scientific, 1998). But that book still falls short
in showing students suitable case studies that are relevant to modern wireless
networks.
On the other hand, most other books addressing network optimization follow
a problem-oriented approach (vs. our method-oriented approach). They do not
offer a systematic treatment of the underlying optimization techniques like we
do in this book. To make this point clear, we quote the following text from the
preface of the book Combinatorial Optimization in Communication Networks,
edited by Maggie Xiaoyan Cheng, Yingshu Li, and Ding-Zhu Du (Springer,
2006), to explain why the problem-oriented approach was adopted by most
authors:
Two approaches were considered: optimization method oriented (starting from combi-
natorial optimization methods and finding appropriate network problems as examples)
and network problem oriented (focusing on specific network problems and seeking
appropriate combinatorial optimization methods to solve them). We finally decided to
use the problem-oriented approach, mainly because of the availability of papers: most
papers in the recent literature appear to address very specific network problems, and
combinatorial optimization comes as a convenient problem solver.
Such a problem-oriented approach offers a convenient way of composing a

book quickly (i.e., by assembling some research papers in the literature into
an edited volume). But books based on such a problem-oriented approach,
although useful as a reference book, do not teach graduate students optimiza-
tion techniques in a systematic manner. This critical dearth in the existing lit-
erature was our main motivation for writing this book and bringing it to the
community.
Acknowledgments
This book is the fruit of close collaboration among the three authors for more
than ten years. We would like first to thank the former and current members of
our research group. In particular, many thanks to: Jia (Kevin) Liu, whose work
led to Chapter 3, Sushant Sharma, whose work led to Chapter 4, Liguang Xie,
whose work led to Chapter 7, Dr. Shiwen Mao, whose work led to Chapter 11.
We want to thank Huacheng Zeng, Liguang Xie, Xu Yuan, and Canming Jiang
for their help in proofreading some of the chapters. They also contributed to the
preparation and revision of the solution manual and Powerpoint Slides. With-
out their help, this book would not have reached its current shape. Some other
former and current members of our group, whose names were not mentioned
above but who contributed to this book in many other ways, include Sastry
Kompella, Cunhao Gao, Tong Liu, Xiaojun Wang, Xiaolin Cheng, Dr. Rongbo
Zhu, Dr. Lili Zhang, and Dr. Wangdong Qi.
We also want to thank the students in our ECE/CS 6570 class (Advanced
Foundations of Networking) over the years, who offered valuable feedback
to different versions of this book and helped us gauge the best match of such
materials for a graduate course in networking. In particular, those students who
took ECE/CS 6570 in Fall 2012 directly contributed to proofreading the final
book manuscript and their feedback is greatly appreciated.
We want to thank Dr. Philip Meyler, Acquisitions Editor of Cambridge Uni-
versity Press, who showed a genuine interest in the initial conception of this
book and encouraged us to move forward for a book proposal. He has also
been extremely patient with us when we requested a one year extension for
the final delivery of our manuscript. We thank him for his trust, patience, and
understanding, which allowed us to work on our schedule to bring this book
to reality. Looking back, we feel really lucky that we chose the best publisher
for this book. During the manuscript preparation stage, we worked with three
different Assistant Editors of Cambridge University Press Elizabeth Horne,
Kirsten Bot, and Sarah Marsh. We thank all three of them, who worked dili-
gently with us at each step along the way to make this book a polished product.
xv Acknowledgments
Tom Hou would like to thank Scott Midkiff, who recruited him to join the
Electrical and Computer Engineering Department at Virginia Tech in 2002.
Over the years, Scott has been a great colleague, a close friend, a resourceful
mentor, and, most recently, a supportive department head. The environment
that Scott and the department were able to offer has been instrumental to Toms
success in research and scholarship.
Finally, we would like to thank the National Science Foundation (NSF) and
the Office of Naval Research (ONR), whose funding support of our research
over the years led to this book.
Copyright permissions
A highlight of this book is to include, in each chapter, a comprehensive presen-

tation of a case study that demonstrates how the particular optimization method
can be applied in solving a wireless networking problem. These case studies
are based on a number of papers written by the authors. The following list
acknowledges these publications and their respective journals and publishers
on a chapter-by-chapter basis. Portions of these papers have been adapted with
permission from the publishers as required. All rights are reserved as stipulated
by the various copyright agreements.
Chapter 2
Y.T. Hou, Y. Shi, and H.D. Sherali, Rate allocation and network lifetime prob-
lems for wireless sensor networks, IEEE/ACM Transactions on Networking,
vol. 16, no. 2, pp. 321334, April 2008. Copyright 2008 by, and with kind
permission from, IEEE.
Chapter 3
J. Liu, Y.T. Hou, Y. Shi, and H.D. Sherali, Cross-layer optimization for
MIMO-based wireless ad hoc networks: routing, power allocation, and band-
width allocation, IEEE Journal on Selected Areas in Communications, vol. 26,
no. 6, pp. 913926, August 2008. Copyright 2008 by, and with kind permis-
sion from, IEEE.
Chapter 4
S. Sharma, Y. Shi, Y.T. Hou, and S. Kompella, An optimal algorithm for relay
node assignment in cooperative ad hoc networks, IEEE/ACM Transactions on
Networking, vol. 19, issue 3, pp. 879892, June 2011. Copyright 2011 by,
and with kind permission from, IEEE.
Chapter 5
Y. Shi, Y.T. Hou, and H. Zhou, Per-node based optimal power control
for multi-hop cognitive radio networks, IEEE Transactions on Wireless
xvii Copyright Permissions
Communications, vol. 8, no. 10, pp. 52905299, October 2009. Copyright

2009 by, and with kind permission from, IEEE.
Chapter 6
Y. Shi, Y.T. Hou, S. Kompella, and H.D. Sherali, Maximizing capacity in mul-
tihop cognitive radio networks under the SINR model, IEEE Transactions on
Mobile Computing, vol. 10, no. 7, pp. 954967, July 2011. Copyright 2011
by, and with kind permission from, IEEE.
Chapter 7
L. Xie, Y. Shi, Y.T. Hou, and H.D. Sherali, Making sensor networks immor-
tal: an energy-renewal approach with wireless power transfer, IEEE/ACM
Transactions on Networking, vol. 20, issue 6, pp. 17481761, December 2012.
Copyright 2012 by, and with kind permission from, IEEE.
Chapter 8
Y. Shi and Y.T. Hou, Optimal base station placement in wireless sensor net-
works, ACM Transactions on Sensor Networks, vol. 5, issue 4, article 32,
November 2009. Copyright 2009 by, and with kind permission from, ACM.
Chapter 9
Y. Shi and Y.T. Hou, Some fundamental results on base station movement
problem for wireless sensor networks, IEEE/ACM Transactions on Network-
ing, vol. 20, issue 4, pp. 10541067, August 2012. Copyright 2012 by, and
with kind permission from, IEEE.
Chapter 10
Y.T. Hou, Y. Shi, and H.D. Sherali, Spectrum sharing for multi-hop network-
ing with cognitive radios, IEEE Journal on Selected Areas in Communica-
tions, vol. 26, no. 1, pp. 146155, January 2008. Copyright 2008 by, and
with kind permission from, IEEE.
Chapter 11
S. Mao, Y.T. Hou, X. Cheng, H.D. Sherali, S.F. Midkiff, and Y.-Q. Zhang,
On routing for multiple description video over wireless ad hoc networks,
IEEE Transactions on Multimedia, vol. 8, no. 5, pp. 10631074, October 2006.
Copyright 2006 by, and with kind permission from, IEEE.
CHAPTER
1 Introduction
The present moment is the only moment available to us, and it is the door to
all moments.
Thich Nhat Hanh
1.1 Book overview
The goal of this book is to offer a course to graduate students by demon-

strating an important set of mathematical tools (with a focus on optimization
techniques) and to show how they can be used to address some challenging
problems in wireless networks.
Book organization This book consists of four parts.
Part I, consisting of three chapters, is devoted to optimization and designing

algorithms that can offer optimal solutions.
Part II, consisting of five chapters, is devoted to techniques that can offer
provably near-optimal solutions.
Part III, consisting of two chapters, is devoted to some highly effective
heuristics.
Part IV, consisting of only one chapter, is devoted to some miscellaneous
topics in the broader context of wireless network optimizations. This part
will be expanded in a future edition.
Structure of each chapter Each chapter starts with a brief overview of a

particular optimization technique and subsequently is followed by a compre-
hensive coverage of a case study in wireless networking. The goal of giving a
pointer to the underlying theory at the beginning of each chapter is to offer a
direction to students on what they should explore further in formal course work
or textbooks in these areas. These pointers are not meant to be comprehensive
tutorials, each of which could constitute a book on its own. The first section of
1
2 Introduction
each chapter is not meant to be a replacement of formal courses in operations

research (OR) or computer science (CS) and certainly does not offer a short
cut for students in their formal education in these subject areas.
Instead, the focus of each chapter is on how to apply the technique under dis-
cussion to solve a challenging problem in wireless networking. Each chapter
self-contained and shows all the details involved in problem formulation, refor-
mulation, and customization of optimization techniques in order to devise a
final solution. Our guiding principles in the choice of a case study in each
chapter are the following:
To reflect state-of-the-art wireless network research rather than discuss some

classic but outdated problems that are no longer of current research interest.
In this spirit, the problems that we chose are mainly in the context of multi-
hop wireless networks, with the underlying wireless technologies being cog-
nitive radio (CR), multiple-input multiple-output (MIMO), and cooperative
communications (CC), among others.
To offer a reasonable level of difficulty or challenge in each case study rather
than a simple problem that the students would hardly encounter in their
research. This is in contrast to simple and diverse examples that are typically
presented in standard optimization or algorithmic textbooks. In this book,
we want to show readers how the method introduced in each chapter can be
applied to solve a hard wireless networking problem, which they are likely to
encounter in research. For this purpose, each chapter is dominated by a case
study in terms of length coverage. We believe this approach will help readers
better appreciate the underlying method and to gain a better understanding
on its application in practice.
To offer only essential background on the underlying wireless communi-
cation technology that is needed in formulating the problem rather than a
comprehensive overview of the technology. This is because the main goal
of this book is to learn various optimization techniques and apply each one
to solve a wireless networking problem as a case study rather than offering
a comprehensive tutorial on various wireless communication technologies.
Therefore, we decided to minimize the coverage on wireless technologies
and offer references that the readers can study further on their own.
To help readers develop strong problem formulation and reformulation skills.
This can only be taught with examples with sufficient sophistication and
complexity. We believe such formulation/reformulation skills are important
for research and thus want to teach the readers such skills in detail in each
case study.
Key characteristics of the book
Presents a collection of useful optimization techniques (one technique in

each chapter), with an emphasis on how each technique is put into action to
solve challenging wireless networking problems.
3 1.2 Book outline
Combines techniques from both operations research (OR) and computer sci-
ence (CS) disciplines, with a strong focus on solving optimization problems
in wireless networks.
Shows various tricks and step-by-step details on how to develop optimiza-
tion models and reformulations, particularly in the context of cross-layer
optimization problems involving flow routing (network layer), scheduling
(link layer), and power control (physical layer).
Discusses case studies that focus on multi-hop wireless networks (e.g., ad
hoc and sensor networks) and incorporates a number of advanced physical
layer technologies such as MIMO, cognitive radio (CR), and cooperative
communications (CC).
Contains problem sets at the end of each chapter. PowerPoint slides for each
chapter are available to both the students and instructors. A solutions manual
is available to the instructors.
1.2 Book outline
This book has four parts. Part I of this book, consisting of Chapters 2 to 4,
is devoted to optimization and designing algorithms that can offer optimal
solutions.
Chapter 2 reviews linear programming (LP) and shows how it can be

employed to solve certain problems in wireless networks. Although the LP
methodology itself is rather basic and straightforward, special care is still
needed to ensure that it is used correctly, as we demonstrate in the case study
in this chapter. The case study is rather interesting as it shows that even in
LP-based problem formulations, deep insights can be gained once we dig
deep into it. In particular, the case study considers lexicographic max-min
(LMM) rate allocation and node lifetime problems in a wireless sensor net-
work (WSN). We introduce the parametric analysis (PA) technique, which is
very useful in its own right. The concept of LMM is also important, and can
be employed as a fairness criterion for other problems in wireless network-
ing research. Through the case study in this chapter, the readers will gain a
rather deep understanding of LP and its applications in wireless networks.
Chapter 3 reviews convex programming, which is a popular and power-
ful tool for studying nonlinear optimization problems. Once a problem is
shown to be a convex program, then there are standard solution techniques
and we may even directly apply a solver to obtain an optimal solution.
For many cross-layer convex optimization problems, the research commu-
nity is more interested in exploring a solution in its dual domain. There
are two reasons for this approach. First, many cross-layer problems, once
properly formulated in the dual domain, can be decomposed into sub-
problems, each of which may be decoupled from the other layers. Such a
4 Introduction
layering-based decomposition offers better insights and interpretations for

the underlying problems. Second, once a problem is decomposed in its dual
domain, the solution may be implemented in a distributed fashion, which is a
highly desirable feature for networking researchers. In the case study of this
chapter, we study a cross-layer optimization problem for a multi-hop MIMO
network, which involves variables at the transport, network, link, and physi-
cal layers. We show that the problem can be formulated as a convex program.
By studying the problem in its dual domain, we show that the dual problem
can be decomposed into two subproblems: one subproblem solely involving
variables at the transport and network layers and the other problem involving
variables at the link and physical layers. We describe how the dual problem
can be solved by a cutting-plane method and how the solution to the primal
problem can be recovered from the solution to the dual problem.
Chapter 4 illustrates how an optimization problem can be solved by clever
algorithmic techniques from CS. For certain problems, general optimization
methods from OR may not always be the best approach. In fact, a formu-
lation following ORs optimization approach may lead to a solution with
nonpolynomial-time complexity. But a solution with nonpolynomial-time
complexity does not mean that the problem is not in P. In fact, we may
well develop a different algorithm to solve the problem with polynomial-
time complexity. This is what we illustrate in this chapter, where we develop
a polynomial-time algorithm. In the case study, we consider a relay node
assignment problem in cooperative communications (CC). Our objective is
to assign a set of available relay nodes to different sourcedestination pairs
so as to maximize the minimum data rate among all the pairs. Following
the OR optimization approach, we show that the problem can be formulated
as a mixed-integer linear programming (MILP) problem, which is NP-hard
in general. But this does not mean that the problem is NP-hard. Instead, by
following a CS algorithm design approach, we develop a polynomial-time
exact algorithm for this problem. A novel idea in this algorithm is a linear
marking mechanism, which is able to achieve polynomial-time complexity
at each iteration. We give a formal proof of the optimality of the algorithm.
Part II of this book, consisting of Chapters 5 to 9, is devoted to techniques

that can offer provably near-optimal solutions.
Chapter 5 presents the branch-and-bound framework and shows how it can

be applied to solve discrete and combinatorial optimization problems. Such
problems are typically considered most difficult in nonconvex optimization
and the branch-and-bound framework offers a general purpose and effec-
tive approach. The effectiveness of branch-and-bound resides in the care-
ful design of each component of its framework, such as computation of a
lower bound, local search of an upper bound, and selection of partitioning
variables (in the case of a minimization problem). It should be noted that
5 1.2 Book outline
the worst-case complexity of a branch-and-bound-based method remains

exponential, although a judicious design of each component could achieve
reasonable computational times in practice. In the case study, we consider a
per-node power control problem for a multi-hop CRN. This problem has a
large design space that involves a tight coupling relationship among power
control, scheduling, and flow routing, which is typical for a cross-layer opti-
mization problem. We develop a mathematical model and a problem formu-
lation, which is a mixed-integer nonlinear programming (MINLP) problem.
We show how to apply the branch-and-bound framework to design a solution
procedure.
Chapter 6 presents the reformulation-linearization technique (RLT) for
deriving tight linear relaxations for any monomial. Simply put, RLT can be

applied to any polynomial term of the form ni=1 (xi )ci in variables xi , where
the ci -exponents are constant integers. Given such generality, RLT is a pow-
erful tool in deriving tight linear relaxations. In the case study, we consider
a throughput maximization problem in a multi-hop CRN under the signal-
to-interference-and-noise-ratio (SINR) model. We develop a mathematical
formulation for joint optimization of power control, scheduling, and flow
routing. We present a solution procedure based on the branch-and-bound
framework and apply RLT to derive tight linear relaxations for a product of
variables. In this case study, we also learn how to identify the core optimiza-
tion space for the underlying problem and how to exploit different physical
interpretations of the core variables in developing a solution.
Chapter 7 presents a linear approximation algorithm, which is a powerful
method to tackle certain nonlinear optimization problems. We show how
such an approach could be employed to solve a nonlinear programming
(NLP) problem in a wireless sensor network (WSN). In addition to the linear
approximation technique, the problem in the case study is interesting on its
own, and shows how the so-called wireless energy transfer technology can
be employed to address network lifetime problems in a WSN.
Chapter 8 shows how to design a polynomial-time approximation algorithm
to provide an (1 )-optimal solution to a nonconvex optimization problem.
The case study focuses on a classic base station placement problem in a
WSN. The design of the (1 )-optimal approximation algorithm is based
on several clever techniques such as discretization of cost parameters (and
distances), partitioning of the search space into a finite number of subareas,
and representation of subareas with fictitious points (with tight bounds on
costs). These three techniques can be exploited to develop approximation
algorithms for other problems. We prove that the approximation algorithm is
(1 )-optimal.
Chapter 9 is a sequel to Chapter 8. Again, our interest is on the design of a
(1 )-optimal approximation algorithm for a mobile base station problem.
But the problem is much harder than that in the last chapter. By allowing
the base station to be mobile, both the location of the base station and the
6 Introduction
multi-hop flow routing in the network are time-dependent. To address this

problem, we show that as far as the network lifetime objective is con-
cerned, we can transform the time-dependent problem to a location (space)-
dependent problem. In particular, we show that flow routing only depends
on the base station location, regardless of when the base station visits this
location. Further, the specific time instances for the base station to visit a
location are not important, as long as the total sojourn time for the base
station to be present at this location is the same. This result allows us to
focus on solving a location-dependent problem. Based on the above result,
we further show that to obtain a (1 )-optimal solution to the location-
dependent problem, we only need to consider a finite set of points within
the smallest enclosing disk for the mobile base stations location. Here, we
follow the same approach as that in Chapter 8, i.e., discretization of energy
cost through a geometric sequence, division of a disk into a finite number
of subareas, and representation of each subarea with a fictitious cost point
(FCP). Then we can find the optimal sojourn time for the base station to stay
at each FCP (as well as the corresponding flow routing solution) so that the
overall network lifetime (i.e., sum of the sojourn times) is maximized via a
single LP Problem. We prove that the proposed solution can guarantee that
the achieved network lifetime is at least (1 ) of the maximum (unknown)
network lifetime. This chapter offers some excellent examples on how to
transform a problem from time domain to space domain and how to prove
results through construction.
Part III, consisting of Chapters 10 and 11, is devoted to some highly effective
heuristics.
Chapter 10 presents an effective approach to address a class of mixed-

integer optimization problems. The technique, called sequential fixing (SF),
is designed to iteratively determine (fix) binary integer variables. It is a
heuristic procedure and has polynomial-time complexity. Its performance
is typically measured by comparing its solution value to some performance
bound, e.g., a lower bound for a minimization problem, or an upper bound
for a maximization problem. Based on our own experience, we find that that
this SF technique is very efficient and can offer highly competitive solutions.
As a case study, we study an optimization problem in a multihop CRN. Since
the problem formulation is an MINLP problem, we develop a lower bound to
estimate the optimal objective value. Subsequently, we present an SF algo-
rithm for this optimization problem. Numerical examples show that the solu-
tions produced by this SF algorithm can offer objective values that are very
close to the computed lower bounds, thus confirming their near-optimality.
Chapter 11 presents metaheuristic methods, which are an important class of
heuristic methods and have been applied to solve some very complex prob-
lems in wireless networks. In this chapter, we give a review of some well-
7 1.3 How to use this book
known metaheuristic methods (e.g., basic local search, simulated annealing

(SA), tabu search (TS), and genetic algorithms (GA)). In the case study,
we focus on developing a GA-based method to solve a multi-path routing
problem for MD video. We find that a GA-based solution is eminently
suitable to address this particular problem, which involves complex objec-
tive functions and exponential solution space. By exploiting the survival-of-
the-fittest principle, a GA-based solution is able to evolve to a population
of better solutions after each iteration and eventually offers a near-optimal
solution.
Part IV, currently consisting of only Chapter 12, is devoted to some miscel-
laneous topics in the broader context of wireless network optimizations. This
part will be further expanded to include other topics in a future book edition.
Chapter 12 presents an asymptotic capacity analysis for wireless ad hoc net-

works. Such an analysis addresses an achievable per-node throughput when
the number of nodes goes to infinity. We focus on so-called random net-
works, where each node is randomly deployed and each node has a randomly
chosen destination node. In this asymptotic capacity analysis, the results are
derived in the form of (), O(), and () and the underlying analysis is
very different from what we do in the other chapters, which focus on opti-
mization problems for finite-sized networks. We show that the asymptotic
capacity analysis heavily depends on the underlying interference model. In
this chapter, we consider three interference models (i.e., the protocol model,
the physical model, and the generalized physical model) and show how to
develop asymptotic capacity bounds for each model.
1.3 How to use this book
This book is written as a textbook and is mainly aimed at graduate stu-

dents (particularly students in electrical and computer engineering) pursuing
advanced research and study in wireless networks. The book could be adopted
for a second or third graduate course in networking. The prerequisites for this
course are a graduate course in networking.
At Virginia Tech, we have been this book in an Electrical and Computer
Engineering (ECE) and Computer Science (CS) cross-listed course titled
ECE/CS 6570: Advanced Foundations of Networking. The prerequisite for
this course is a graduate course in networking. We cover Chapters 2, 3, 4, 5,
8, 10, and 11 in a one semester course and the response from the students has
been overwhelmingly positive! For each chapter, we have prepared PowerPoint
slides, which are available to both the students and instructors. We have also
8 Introduction
prepared a solutions manual for all end-of-chapter problems, which is available

to the instructors.
In addition to its primary role as a graduate textbook, researchers in
academia who are active in conducting research in wireless networks will
find this book a very useful reference to expand their toolboxes in problem
solving. Further, researchers and engineers in industry and government labo-
ratories who perform active research in wireless networks will also find this
book to be a useful reference.
PART
I Methods for Optimal Solutions

CHAPTER
2 Linear programming and

applications
The real voyage of discovery consists not in seeking new landscapes but in
having new eyes.
Marcel Proust
2.1 Review of key results in linear programming
Linear programming (LP) is a problem consisting of a linear objective function

and a set of linear constraints (equations or inequalities) with real variables.
Such a problem aims to maximize or minimize a specific objective function
by systematically choosing the values of the real variables within an allowed
set (i.e., solution space). In an LP problem, both the objective function to be
optimized and all the constraints restricting the variables are linear.
A general form of an LP problem is as follows:
Maximize c1 x1 + c2 x2 + . . . + cn xn
subject to a11 x1 + a12 x2 + . . . + a1n xn b1
a21 x1 + a22 x2 + . . . + a2n xn b2
a31 x1 + a32 x2 + . . . + a3n xn = b3
..
.
am1 x1 + am2 x2 + . . . + amn xn = bm
xj 0 (1 j n).
The function c1 x1 + c2 x2 + . . . + cn xn is the objective function to be maxi-
mized, where c1 , c2 , . . . , cn are constant coefficients, and x1 , x2 , . . . , xn are
the so-called decision variables to be determined. The equality (or inequal-

ity) nj=1 aij xj = (or or ) bi is the ith constraint, where aij is a constant
11
12 Linear programming and applications
coefficient. The constant column vector with the ith component bi , 1 i m,

is the right-hand-side (RHS) vector. The restrictions xj 0, 1 j n are
called nonnegativity constraints. Sometimes, they may be replaced by general
lower and upper bounds on the decision variables. Observe that a maximization
problem can be easily converted into a minimization problem and conversely,
noting that
Maximum c1 x1 + c2 x2 + . . . + cn xn
= Minimum c1 x1 c2 x2 . . . cn xn .
In a general form of an LP problem, both equalities and inequalities may
appear. Performing a simple manipulation on the inequalities (by adding or
subtracting nonnegative slack variables, see [10] for more details), any LP
problem can be transformed to an equivalent standard form as follows:
Maximize cT x
subject to Ax = b (2.1)
x 0,
where x is the vector of nonnegative continuous variables, c is the vector of
coefficients in the objective function, A is the matrix of coefficients in the con-
straints, and b is the RHS vector for the constraints. This simple transformation
is convenient for developing a general algorithmic procedure to solve and ana-
lyze LPs [10].
An LP problem can be solved optimally in a time complexity of O(nA 3 )
(e.g., using interior point methods [10]), where nA is the number of variables
in the standard form representation. Therefore, once we are able to formu-
late a problem as an LP, we can solve this problem in polynomial time by
employing open-source solvers (e.g., GLPK [52]) or commercial solvers (e.g.,
CPLEX [31] and Lindo [95]).
For an LP problem in standard form, there is another associated LP problem
called the dual problem, which is given as follows:
Minimize wT b
subject to AT w c,
where the vector of dual variables w is unrestricted in sign [10]. The dual
problem has some important properties and economic interpretation, and can
also be used to obtain the solution to the original LP problem (e.g., with the
dual simplex method or the primal-dual algorithm [10]). Based on the duality
relationship in LP theory, both the original LP and its dual LP can be solved
simultaneously by standard LP techniques in polynomial time.
We want to emphasize that all the variables in an LP problem should be real
(continuous) variables. If an optimization problem has a linear objective func-
tion, linear constraints, and some real variables and some integer/binary vari-
ables, then this optimization problem is a mixed-integer linear programming
13 2.2 Case study
(MILP), which is not an LP problem. A special case of MILPs are integer

linear programs (ILP), where all the decision variables are integer/binary
restricted. Unlike LP, there does not exist a polynomial-time algorithm to solve
an MILP or an ILP optimally.
In the rest of this chapter, we offer a case study on how LP can be applied to
address some interesting but difficult problems from wireless sensor networks.
2.2 Case study: Lexicographic max-min rate allocation and

node lifetime problems
Consider a wireless sensor network consisting of battery-powered nodes for
data collection. Although there have been significant improvements in pro-
cessor design and computing, advances in battery technology still lag behind,
making energy resource considerations the fundamental challenge in wire-
less sensor networks. Consequently, there have been active research efforts
on performance limits of wireless sensor networks. These performance limits
include, among others, network capacity (e.g., [78]) and network lifetime (e.g.,
[24]). Network capacity typically refers to the maximum amount of bit volume
that can be successfully delivered to the base station (sink node) by all the
nodes in the network, while network lifetime refers to the maximum time limit
for which all the nodes in the network remain alive until one or more nodes
drain up their energy.
In this chapter, we consider an overarching problem that encompasses both
performance metrics. In particular, we study the network capacity problem
under a given network lifetime requirement. Specifically, for a wireless sensor
network where each node is provisioned with an initial energy, if all nodes are
required to live up to a certain lifetime, what is the maximum amount of bit
volume that can be generated by the entire network? At first glance, it appears
desirable to maximize the sum of rates for all the nodes in the network, sub-
ject to the condition that each node can meet the network lifetime requirement.
Mathematically, this problem can be formulated as a linear programming (LP)
problem (see Section 2.3.2) within which the objective function is defined as
the sum of rates over all the nodes in the network and the constraints require
that (i) flow balance is preserved at each node, and (ii) the energy constraint
at each node is met for the given network lifetime requirement. However, the
solution to this problem shows (see Section 2.7) that there exists a severe bias
in rate allocation among the nodes, despite that the sum of bit rates is maxi-
mized. In particular, those nodes that consume the least amount of power on
their data path toward the base station are allocated with much more bit rates
than other nodes in the network. Consequently, the data collection behavior for
the entire network only favors certain nodes that have this property, while other
nodes will be unfavorably penalized with much smaller bit rates.
The fairness issue associated with the network capacity maximization objec-
tive calls for careful consideration in rate allocation among the nodes. In this
chapter, we study the rate allocation problem in an energy-constrained sen-
sor network for a given network lifetime requirement. Our objective is to
achieve a certain measure of optimality in the rate allocation that takes into
account both fairness and bit rate maximization. We employ the so-called Lex-
icographic Max-Min (LMM) criterion [99], which maximizes the bit rates
for all the nodes under the given energy constraint and network lifetime
requirement. At the first level, the smallest rate among all the nodes is max-
imized. Then, we continue to maximize the second smallest rate level and so
forth. The LMM rate allocation criterion is appealing since it addresses both
fairness and efficiency (i.e., bit rate maximization) in an energy-constrained
network.
A naive approach to the LMM rate allocation problem would be to apply
a max-min-like iterative procedure. Under this approach, successive LPs are
employed to calculate the maximum rate at each level based on the avail-
able energy for the remaining nodes, until all nodes use up their energy. We
call this approach serial LP with energy reservation (SLP-ER). We will
show that, although SLP appears intuitive, it is likely to offer an incorrect
solution. To understand how this could happen, we must understand a funda-
mental difference between the LMM rate allocation problem described here
and the classical max-min rate allocation in [14]. Under the LMM rate allo-
cation problem, the rate allocation is implicitly coupled with a flow routing
problem, while under the classical max-min rate allocation, there is no rout-
ing problem involved since the routes for all flows are given. As it turns out,
for the LMM rate allocation problem, any iterative rate allocation approach
that requires energy reservation at each iteration is incorrect. This is because,
unlike max-min, which addresses only the rate allocation problem with fixed
routes and yields a unique solution at each iteration, for the LMM rate allo-
cation problem, there usually exist nonunique flow routing solutions corre-
sponding to the same rate allocation at each level. Each of these flow routing
solutions will yield different available energy levels on the remaining nodes
for future iterations and so forth. This will lead to a different rate alloca-
tion vector, which may not coincide with the optimal LMM rate allocation
vector.
In this chapter, we show a correct approach to solve the LMM rate alloca-
tion problem. Our approach employs the so-called parametric analysis (PA)
technique [10] in LP to determine the minimum set of nodes at each rate level
that must deplete their energy. We call this approach serial LP with PA (SLP-
PA). We also extend the PA technique for the LMM rate allocation problem to
address the so-called maximum node lifetime curve problem in [20], which
we call the LMM node lifetime problem. More importantly, we show that
there exists a simple and elegant mirror relationship between the LMM rate
15 2.3 System modeling and problem formulation
allocation problem and the LMM node lifetime problem. As a result, it is suf-
ficient to solve only one of these two problems.
The remainder of this chapter is organized as follows. Section 2.3 intro-
duces the network and energy model, and formulates the LMM rate alloca-
tion problem. Section 2.4 describes the SLP-PA algorithm for the LMM rate
allocation problem. Section 2.5 introduces the LMM node lifetime problem
and shows how to apply the SLP-PA algorithm to solve it. Section 2.6 reveals
an interesting mirror relationship between the LMM rate allocation problem
and the LMM node lifetime problem. Section 2.7 presents numerical results.
Section 2.8 summarizes this chapter.
2.3 System modeling and problem formulation
We consider a two-tier architecture for wireless sensor networks. Figs. 2.1(a)

and (b) show the physical and hierarchical network topologies for such a net-
work, respectively. There are three types of nodes in the network namely,
micro-sensor nodes (MSNs), aggregation and forwarding nodes (AFNs) and
a base station (BS). The MSNs can be application-specific sensor nodes (e.g.,
temperature, chemical, video) and they constitute the lower-tier data collection
network. They are deployed in groups (or clusters) at strategic locations for
surveillance and monitoring applications. The MSNs are small and low cost.
The objective of an MSN is simple: once triggered by an event, it starts to
collect sensing data and sends it to the local AFN (in one hop).
For each cluster of MSNs, there is one AFN, which is different from an MSN
in terms of physical properties and functions. The primary functions of an AFN
are: (1) data aggregation (or fusion) for data flows from the local cluster of
MSNs, and (2) forwarding (or relaying) the aggregated information to the next
hop AFN (ultimately to the base station). Although an AFN is expected to be
provisioned with much more energy than an MSN, it also consumes energy
at a substantially higher rate (due to data transmission over longer distances).
Consequently, an AFN has a limited lifetime. Upon depletion of energy at an
AFN, we expect that the coverage for the particular area under surveillance is
lost, despite the fact that some of the MSNs within the cluster may still have
remaining energy.
The third component in the two-tier architecture is the base station. The
base station is, essentially, the sink node for data streams from all the AFNs
in the network. We assume that there is sufficient energy resource available
at the base station and thus there is no energy constraint at the base station.
In summary, from networking perspective, the lower-tier MSNs are for data
acquisition, while the upper-tier AFNs are for transporting data to the base
station.
Figure 2.1 Reference

architecture for two-tier wireless
sensor networks.
Base station
(BS)
Aggregation and forwarding node

(AFN) Micro-sensor node
(MSN)
(a) Physical topology.
Base station
(BS)
Upper tier
AFN
Lower tier
MSN
(b) A hierarchical view.
2.3.1 Energy model

Table 2.1 lists the notation used in this chapter. We focus on energy consump-
tion at AFNs. For AFN i, we assume that the aggregated bit rate collected
locally (after data fusion) is ri , i = 1, 2, . . . , N . These collected local bit
streams must be relayed to the base station. Our objective is to maximize the
ri -values according to the LMM criterion (see Definition 2.1) under a given
network lifetime requirement.
For an AFN, we assume that energy consumption due to transmission and
reception is the dominant source of energy consumption [4]. The power dissi-
pation at a radio transmitter can be modeled as
utij = Cij fij , (2.2)

Table 2.1 Notation.
General notation for the LMM rate and LMM lifetime problems
B The base station

Cij (or CiB ) The power consumption coefficient for transmitting data from AFN
i to AFN j (or the base station B)
ei The initial energy at AFN i
fij (or fiB ) Data rate from AFN i to AFN j (or the base station B)
n The number of distinct elements in the sorted LMM-optimal
rate/lifetime vector
N Total number of AFNs in the network
Si The minimum set of nodes that reach their energy constraint limits at
the ith level
Si The set of all possible AFNs that may reach their energy constraint
limits at the ith level, Si Si
Vij (or ViB ) Total data volume from AFN i to AFN j (or the base station B)
Path-loss index
1 , 2 Constant terms in transmission power model
The power consumption coefficient for receiving data
Symbols used for the LMM rate problem
ri The local bit rate collected at AFN i
gi The ith element in the sorted LMM-optimal rate vector, where g1
g2 . . . gN
T The network lifetime requirement
i The ith rate level in the sorted LMM-optimal rate vector, i.e.,
1 (= g1 ) < 2 < . . . < n (= gN )
i = i i1 , the difference between i and i1
Symbols used for the LMM lifetime problem
ri The rate requirement at AFN i
ti The node lifetime at AFN i
i The ith element in the sorted LMM-optimal lifetime vector, where
1 2 . . . N
i The ith drop point in the sorted LMM-optimal lifetime vector, i.e.,
1 (= 1 ) < 2 < . . . < n (= N )
i = i i1 , the difference between i and i1
where utij is the power dissipated at AFN i when it is transmitting to node j ,

fij is the rate from AFN i to node j , Cij is the power consumption cost of
radio link i j and is given by
Cij = 1 + 2 dij , (2.3)
where 1 and 2 are two constant terms, dij is the distance between these
two nodes, and is the path-loss index. Typical values for these param-
eters are 1 = 50 nJ/b, 2 = 0.0013 pJ/b/m4 [67], and = 4 [129]. Since
the power level of an AFNs transmitter can be used to control the distance
coverage of an AFN (e.g., [127; 131; 166]), different network flow rout-
ing topologies can be formed by adjusting the power level of each AFNs
transmitter.
The power dissipation at a receiver can be modeled as

uri = fki , (2.4)
k=i

where k=i fki (in b/s) is the rate of the received data stream at AFN i. A
typical value for the parameter is 50 nJ/b [67].
To ensure data transmissions at different AFNs do not interfere with each
other, we assume that a scheduling mechanism has been employed. The details
of such a scheduling mechanism is beyond the scope of this chapter.
2.3.2 The LMM rate allocation problem

Before we formulate the LMM rate allocation problem, we revisit the maxi-
mum capacity problem (with bias in rate allocation) that was discussed in
Section 2.2. For a network with N AFNs, suppose that the rate of AFN i is
ri , and that the initial energy at this node is ei (i = 1, 2, . . . , N). For a given
network lifetime requirement T (i.e., each AFN must remain alive for at least
the time duration T ), the maximum sum of rates can be found by the following
LP problem:
N
MaxCap Maximize i=1 ri

subject to fiB + j =i fij k=i fki = ri (1 i N ) (2.5)

k=i fki T + j =i Cij fij T + CiB fiB T ei (1 i N ) (2.6)
fij , fiB 0 (1 i, j N, j = i).
The set of constraints in (2).5 are the flow balance equations. They state that
the total bit rate transmitted by AFN i is equal to the total bit rate received by
AFN i from other AFNs, plus the bit rate generated locally at this node (ri ).
Note that we allow flow splitting at a node so as to achieve more flexibility
in flow routing and load balancing in the network. The set of constraints in
(2).6 are the energy constraints. They state that for a given network lifetime
requirement T , the energy consumed for communications (i.e., transmitting
and receiving) cannot exceed the initial available energy.
Note that fki , fij , fiB , and ri are variables and T is a constant (the
given network lifetime requirement). Since MaxCap is an LP problem, it
can be solved in polynomial time [10]. Unfortunately, as we will see in the
numerical results (Section 2.7), the solution to this MaxCap problem favors
those AFNs whose data paths consume the least amount of power toward
the base station. Consequently, although the sum of rates is maximized over
T , the specific bit rate allocation among the AFNs (i.e., ri s) favors those
AFNs that have this property, while the other AFNs are unfavorably allocated
with much smaller (even close to 0) data rates. As a result of this unfair-
ness, the effectiveness of the sensor network in performing data collection is
questionable.
To address this fairness issue, we can employ the so-called lexicographic

max-min (LMM) rate allocation strategy [99], which bears some similarity to
the max-min rate allocation in data networks [14].1 Under LMM rate allo-
cation, we start with the objective of maximizing the bit rate of all nodes
until one or more nodes reach their limits for the given network lifetime
requirement. Given that the first level of the smallest rate allocation is max-
imized, we continue to maximize the second level of rate for the remain-
ing nodes that still have available energy, and so forth. More formally,
denote g = [g1 , g2 , . . . , gN ] as the sorted version (i.e., g1 g2 . . . gN )
of the rate vector r = [r1 , r2 , . . . , rN ], with ri corresponding to the rate
of node i. We then have the following definition for LMM-optimal rate
allocation:
Definition 2.1
LMM-optimal rate allocation For a given network lifetime require-
ment T , a sorted rate vector g = [g1 , g2 , . . . , gN ] yields an LMM-optimal
rate allocation if and only if for any other sorted rate allocation vector
g = [g1 , g2 , . . . , gN ] with g1 g2 . . . gN there exists k, 1 k N ,
such that gi = gi for 1 i k 1 and gk > gk .
2.3.3 Two incorrect approaches

Before we present a correct solution to the optimal LMM rate allocation prob-
lem, we discuss two incorrect solutions and explain why they cannot provide
an LMM-optimal solution.
Serial LP with energy reservation Based on the LMM-optimal definition,
we can calculate the first level optimal rate 1 = g1 easily through the follow-
ing LP problem:
Maximize 1

subject to fiB + j =i fij k=i fki 1 = 0 (1 i N )

k=i Tfki + j =i Cij Tfij + CiB TfiB ei (1 i N )
fij , fiB 0 (1 i, j N, j = i).
Once we obtains a solution with the maximum 1 , we can also calculate the
energy consumption at each node under this flow routing solution. Then we can
check whether or not a node has any remaining energy. If there are some nodes
that still have remaining energy, then we can construct another LP problem
to further increase their data rates (with a maximum rate increment of 2 ).
1 However, there is a significant difference between max-min and LMM, which we will discuss
shortly.
This process terminates until all nodes use up their energy. Since energy is
reserved at each node after each iteration, we call this naive approach Serial
LP with Energy Reservation (SLP-ER).
SLP-ER cannot provide an LMM-optimal solution because there is a funda-
mental difference in the nature of the LMM rate allocation problem described
here and the classical max-min rate allocation problem in [14]. The LMM rate
allocation problem implicitly couples a flow routing problem (i.e., a determi-
nation of the fij and fiB for the entire network), while the classical max-min
rate allocation explicitly assumes that the routes for all the flows are given a
priori and remain fixed. For the LMM rate allocation problem, starting from
the first iteration, there usually exist nonunique flow routing solutions corre-
sponding to the same maximum rate level. Consequently, each of these flow
routing solutions, once chosen, will yield different remaining energy levels on
the nodes for future iterations and so forth. This will lead to a nonunique rate
allocation vector, which may not coincide with the LMM-optimal rate vector
(see numerical examples in Section 2.7).
Serial LP with rate reservation Instead of reserving energy based on a

flow routing solution at each rate level, another approach is to only deter-
mine the set of nodes that use up their energy at this level. The final flow
routing solution can be deferred to the last iteration. Since rate is reserved
during each iteration, we call this approach Serial LP with Rate Reservation
(SLP-RR).
SLP-RR cannot achieve an LMM-optimal rate vector either. This may be
harder to understand. But it is still due to the existence of non-unique flow
routing solutions at each rate level, whereby a node that uses up its energy
in one flow routing solution may still have remaining energy in another flow
routing solution. Therefore, if we reserve rates just based on one flow routing
solution, the set of nodes thought to have used up their energy at this rate level
may include some extra nodes that would otherwise be allocated with a larger
data rate in another flow routing solution. Consequently, a solution obtained
by SLP-RR may not be LMM-optimal.
2.4 A serial LP algorithm based on parametric analysis
In this section, we present a correct polynomial-time algorithm to solve the

LMM rate allocation problem. We first introduce some of the notation used.
Suppose that the rate vector g = [g1 , g2 , . . . , gN ] is LMM-optimal, with g1
g2 . . . gN . Note that the values of these N rates may not be distinct. To
focus on those distinct rate levels, we remove any repetitive elements in this
vector and rewrite it as [1 , 2 , . . . , n ] such that 1 < 2 < . . . < n , where
1 = g1 , n = gN , and n N . Now, for each i , i = 1, 2, . . . , n, denote Si
21 2.4 A serial LP algorithm based on parametric analysis
as the corresponding set of nodes that use up their energy at this rate. Clearly,

we have ni=1 |Si | = |S| = N , where S denotes the set of all N nodes.
The essence to solving the LMM rate allocation problem is to find the correct
values for 1 , 2 , . . . , n and the corresponding sets S1 , S2 , . . . , Sn , respec-
tively. This can be done iteratively. We first determine the rate level 1 and the
corresponding set S1 , then determine the rate level 2 and the corresponding
set S2 , and so on. In Section 2.4.1, we will show how to determine each rate
level and in Section 2.4.2, we will show how to determine the corresponding
node set.
2.4.1 Determining rate levels

Denote 0 = 0 and S0 = . For l = 1, 2, . . . , n, suppose that we have already
determined 0 , 1 , . . . , l1 and the corresponding sets S0 , S1 , . . . , Sl1 .
The rate level l can be found by the following optimization problem:
Maximize l

l1
subject to fiB + j =i fij k=i fki l = l1 (i Sh ) (2.7)
h=0

fiB + j =i fij k=i fki = h (i Sh , 1 h < l) (2.8)

l1
( k=i fki + j =i Cij fij + CiB fiB )T ei (i Sh ) (2.9)
h=0

( k=i fki + j =i Cij fij + CiB fiB )T = ei (i Sh , 1 h < l) (2.10)
fij , fiB 0 (1 i, j N, j = i).
Note that for l = 1, constraints (2).8 and (2).10 do not exist. For 2 l n,
constraints (2).8 and (2).10 are for those nodes that have already reached their
LMM rate allocation during the previous (l 1) iterations. In particular, the
set of constraints in (2).8 say that the sum of in-coming and local data rates
is equal to the out-going data rates for each node with its LMM-optimal rate
h , 1 h < l. The set of constraints in (2).10 say that for those nodes that
have already reached their LMM-optimal rates, the total energy consumed for
communications has reached their initial energy provisioning. On the other
hand, the constraints in (2).7 and (2).9 are for the remaining nodes that have
not yet reached their LMM-optimal rates. Specifically, the set of constraints in
(2).7 state that, for those nodes that have not yet reached their energy constraint
levels, the sum of in-coming and local data rates is equal to the out-going data
rates. Note that the objective function is to maximize the additional rate l for
those nodes. Furthermore, for those nodes, the set of constrains in (2).9 state
that the total energy consumed for communications should be upper bounded
by the initial energy provisioning.
To facilitate our later discussion on mirror results in Section 2.6, we fur-

ther re-formulate the foregoing LP problem. In particular, we multiply both
sides of (2).7 and (2).8 by T (which is a constant) and denote ViB = fiB T ,
Vij = fij T , and Vki = fki T . Intuitively, Vij and ViB represent the bit volume
that is transferred from node i to j and from node i to B, respectively, during
the lifetime T . We obtain the following LP formulation, which can be solved
using standard LP techniques:
LMM-rate Maximize l

l1
subject to ViB + j =i Vij k=i Vki l T = l1 T (i Sh ) (2.11)
h=0

ViB + j =i Vij k=i Vki = h T (i Sh , 1 h < l)

l1
k=i Vki + j =i Cij Vij + CiB ViB ei (i Sh )

h=0

k=i Vki + j =i Cij Vij + CiB ViB = ei (i Sh , 1 h < l)
Vij , ViB 0 (1 i, j N, j = i).
Although a solution to the LMM rate problem gives the optimal solution for
l at iteration l, it remains to determine the minimum set of nodes correspond-
ing to this l , which is the key difficulty in the LMM rate allocation problem.
Fortunately, this problem can be solved by the so-called parametric analysis
(PA) technique in LP [10].
2.4.2 Determining minimum node set for a rate level

For the minimum node set for a rate level, we first need to know whether or
not this set is unique. This is affirmed in the following lemma. Its proof can be
found in [71].
Lemma 2.1
The minimum node set for each rate level under the LMM-optimal rate
allocation is unique.
With this lemma in place, we now discuss how to determine the minimum
node set Sl corresponding to the rate level l . Denote Sl ( = ) as the set of
nodes for which the constraints (2).9 are binding at the lth iteration in LMM
rate, i.e., Sl include all the nodes that achieve equality in (2).9 at iteration l.
Although it is certain that at least one of the nodes in Sl belong to Sl (the
minimum node set for rate l ), some nodes in Sl may still achieve greater rates
under other flow routing solutions. In other words, if |Sl | = 1, then we must
23 2.4 A serial LP algorithm based on parametric analysis
have Sl = Sl otherwise, we must determine the minimum node set Sl ( Sl )

that achieves the LMM-optimal rate allocation.
The PA technique [10] is most suitable to address this problem. The main
idea of PA is to investigate how an infinitesimal perturbation on some param-
eters of the LMM rate problem can affect the objective function. Specifically,
node i belongs to the minimum node set Sl if and only if a small increase in
node is rate leads to a decrease in the objective value. In our problem, we con-
sider a small increase on the right-hand-side (RHS) of (2).11, i.e., changing bi
in the standard form (2).1 to bi + i , where i > 0. The physical meaning is
that node is data rate increased from l + l1 to l + l1 + Ti . We solve the
updated optimization problem with this new requirement on node is data rate.
Node i belongs to the minimum node set Sl if and only if the new l is smaller,
+
i.e., i l (0) < 0. The details of applying PA to determine the minimum node
set Sl can be found in [71].
2.4.3 Optimal flow routing for LMM rate allocation

After we solve the LMM rate allocation problem iteratively using the pro-
cedure described in Sections 2.4.1 and 2.4.2, the corresponding optimal flow
routing can be obtained by dividing the total bit volume on each link (Vij or
ViB ) by T , i.e.,
Vij ViB
fij = and fiB = . (2.12)
T T
Although the LMM-optimal rate allocation is unique, it is important to note
that the corresponding flow routing solution is not unique. This is because
upon the completion of the LMM rate allocation problem (i.e., upon find-
ing [1 , 2 , . . . , n ]), there usually exist nonunique bit volume solutions (Vij
and ViB values) corresponding to the same LMM-optimal rate allocation. This
result is summarized in the following lemma:
Lemma 2.2
The optimal flow routing solution corresponding to the LMM rate alloca-
tion may not be unique.
We use the following example to illustrate the nonuniqueness of the optimal

flow routing solution for an LMM rate allocation.
Example 2.1
Consider an eight-node network with a topology shown in Fig. 2.2. The
base station B is located at the origin (0, 0). There are two groups of nodes,
G1 and G2 , in the network, with each group consisting of four nodes. Group
G1 consists of AFN1 at (100, 0), AFN3 at (0, 100), AFN5 at (100, 0),
and AFN7 at (0, 100), respectively (all in meters); Group G2 consists of
AFN2 at (100, 100), AFN4 at (100, 100), AFN6 at (100, 100), and
AFN8 at (100, 100), respectively. Assume that all nodes have the same
initial energy e. For a network lifetime requirement of T , we can calculate
(via SLP-PA) that the final LMM-optimal rate allocation for all eight nodes
are identical (perfect fairness), i.e., r1 = r2 = . . . = r8 . We denote ri = g
for 1 i 8.
Y (m)
4 39g/62 x 3 x 2
100
23
39g/62 x
101g/ 62
g/
62
62
g/
x
23
5 101g/ 62 101g/ 62 1
X (m)
100 0 100
101g/ 62
39g/62 x
23
62
g/ x
g/
62
23
x 39g/62 x
6 100
7 8
Figure 2.2 A simple example showing that the optimal flow routing to the LMM rate allocation is not
unique. The range of x is 0 x 39g
62 .
Upon the completion of the SLP-PA algorithm, we also obtain an opti-

mal flow routing solution corresponding to this LMM-optimal rate g. This
optimal flow routing solution has the following flows: f21 = f43 = f65 =
f87 = 3962 g, f2B = f4B = f6B = f8B = 62 g, and f1B = f3B = f5B =
23
f7B = 62 g. We now show that the optimal flow routing solution is

101
nonunique. Since the network has a symmetrical property, it can be eas-

ily verified that for any x, 0 x 3962 g, the LMM-optimal rate allocation
can be achieved if the flow routing solution satisfies the following two con-
ditions: (i) each node in G2 (i.e., AFNs 2, 4, 6, and 8) sends a flow of
x and a flow of 39 62 g x to its two neighboring G1 nodes as shown in
Fig. 2.2, and a remaining flow of 23 62 g directly to the base station; and (ii)
each node in G1 (i.e., AFNs 1, 3, 5, and 7) sends a total amount of 101 62 g
to the base station, which includes x and 39 62 g x from its neighboring
nodes, plus g from itself. Clearly, there are infinitely many flow routing
solutions that meet these two conditions, each of which can be shown to
yield the LMM-optimal rate allocation g with the given network lifetime
requirement T .
25 2.5 SLP-PA for the LMM node lifetime problem
2.4.4 Complexity analysis

We now analyze the complexity of the SLP-PA algorithm. First, we consider
the complexity of finding each nodes rate and the total bit volume transmitted
along each link. At each stage, we solve an LP problem with a complexity of
O(nA 3 ) [10], where nA is the number of variables in the formulated LP (in the
standard form). Since the number of variables is O(N 2 ), the complexity of
solving the LP problem is O(N 6 ). After solving an LP problem at each stage,
we need to determine whether or not a node that just reached its energy binding
constraint belongs to the minimum node set for this stage. In [71], we showed
that the complexity of this is O(N 7 ). Hence, the complexity at each stage
is O(N 6 ) + O(N 7 ) = O(N 7 ). Since there are at most N stages, the overall
complexity is O(N 8 ).
The complexity in finding an optimal flow routing is bounded by the num-
ber of radio links in the network, which is O(N 2 ). Hence, the overall com-
plexity is O(N 8 ) + O(N 2 ) = O(N 8 ). Note that the analysis here gives a
worst-case time complexity. In practice, the run-time of SLP-PA is much
faster.
2.4.5 Extension to variable bit rate

So far, we have considered the case that each AFN generates data at a constant
rate. In practice, an AFN node may not always transmit data and may work in
an onoff mode to conserve energy. In this case, it is necessary to construct an
optimal flow routing solution for a variable bit rate source (where the onoff
mode is a special case). In [70], we have developed techniques to construct an
optimal flow routing solution for a variable bit rate, as long as the average rate
is known. Such an average rate corresponds to the constant rate in this chapter.
As a result, the case of the onoff mode (with a known average rate) can also
be handled using the techniques described in [70].
2.5 SLP-PA for the LMM node lifetime problem
In this section, we show that the SLP-PA algorithm can be used to solve the
so-called maximum node lifetime curve problem in [20], which we define as
the LMM node lifetime problem. We also show that the SLP-PA algorithm is
much more efficient than the one proposed in [20].
2.5.1 The LMM node lifetime problem

The LMM node lifetime problem can be described as follows. For a net-
work with N AFNs, with a given local bit rate ri (fixed) and initial energy
ei for AFN i, i = 1, 2, . . . , N, how can we maximize the network lifetime

for all AFNs in the network? In other words, the LMM node lifetime prob-
lem not only considers how to maximize the lifetime until the first AFN
runs out of energy, but also the lifetimes for the rest of the AFNs in the
network.
More formally, denote the lifetime for each AFN i as ti , i = 1, 2, . . . , N.
Note that the ri s are fixed here, while the ti s are the optimization variables,
which are different from those in the LMM rate allocation problem in the pre-
vious section. Denote [1 , 2 , . . . , N ] as the sorted sequence of the ti -values
in nondecreasing order. Then, the LMM-optimal node lifetime can be defined
as follows:
Definition 2.2
LMM-optimal node lifetime A sorted node lifetime vector
[1 , 2 , . . . , N ] with 1 2 . . . N is LMM-optimal if and only if for
any other sorted node lifetime vector [1 , 2 , . . . , N ], with 1 2 . . .
N , there exists a k, 1 k N such that i = i for 1 i k 1 and
k > k .
2.5.2 Solution
It should be clear that, under the LMM-optimal node lifetime objective, we
must maximize the time until a set of nodes use up their energy (which is also
called a drop point in [20]), while minimizing the number of nodes that drain
up their energy at each drop point. We now show that the SLP-PA algorithm
developed for the LMM rate allocation problem can be applied to solve the
LMM node lifetime problem.
Suppose that [1 , 2 , . . . , N ] with 1 2 . . . N is LMM-optimal.
To keep track of distinct node lifetimes (or drop points) in this vector, we
remove all repetitive elements in the vector and rewrite it as [1 , 2 , . . . , n ]
such that 1 < 2 < . . . < n , where 1 = 1 , n = N , and n N . Corre-
sponding to these drop points, denote S1 , S2 , . . . , Sn as the sets of nodes that
drain up their energy at drop points 1 , 2 , . . . , n , respectively. Then |S1 | +
|S2 | + . . . + |Sn | = |S| = N, where S denotes the set of all N AFNs in the net-
work. The problem is to find the LMM-optimal values of 1 , 2 , . . . , n and
the corresponding sets S1 , S2 , . . . , Sn .
Similar to the LMM rate allocation problem, the LMM node lifetime prob-
lem can be formulated as an iterative optimization problem as follows. Denote
0 = 0, S0 = , and l = l l1 . Starting from l = 1, we solve the fol-
lowing LP problem iteratively:
27 2.6 A mirror result
LMM lifetime Maximize l

l1
subject to ViB + Vij Vki l ri = l1 ri (i Sh ) (2.13)
j =i k=i h=0

ViB + Vij Vki = h ri (i Sh , 1 h < l)
j =i k=i

l1
Vki + Cij Vij + CiB ViB ei (i Sh )
k=i j =i h=0

Vki + Cij Vij + CiB ViB = ei (i Sh , 1 h < l)
k=i j =i
Vij , ViB , l 0 (1 i, j N, j = i).
Comparing the above LMM lifetime problem to the LMM rate problem in
Section 2.4.1, we find that they are mathematically identical. The only difference
is that under the LMM lifetime problem, the local bit rates ri s are constants and
the node lifetimes i s are variables, while under the LMM rate problem, the
ri s are variables and the node lifetimes are all identical (T ), i = 1, 2, . . . , N.
Since the mathematical formulation for the two problems are identical, the
SLP-PA algorithm can be applied to solve the LMM node lifetime problem.
The only issue that we need to be concerned about is the optimal flow rout-
ing solution corresponding to the LMM-optimal lifetime vector. The optimal
flow routing solution here is not as simple as that for the LMM rate allocation
problem, which merely involves a simple division (see (2).12). We refer read-
ers to the appendix at the end of this chapter for an O(N 4 ) algorithm to obtain
an optimal flow routing solution for the LMM-optimal lifetime vector. Similar
to Lemma 2.2, the optimal flow routing solution corresponding to the LMM
node lifetime problem may not be unique.
2.6 A mirror result
In this section, we present an elegant result showing that there is a mirror rela-
tionship between the LMM rate allocation problem and the LMM node lifetime
problem. As a result, it is only necessary to solve only one of the two problems
and the results for the other problem can be obtained via simple algebraic cal-
culations.
To start with, we denote PR as the LMM rate allocation problem where we
have N AFNs in the network and all nodes have a common lifetime require-
ment T (a given constant). Denote ri as the LMM-optimal rate allocation for
node i under PR , i = 1, 2, . . . , N. Similarly, we denote PL as the LMM node
lifetime problem where all nodes have the same local bit rate R (constant).
Denote ti as the LMM node lifetime for node i under PL , i = 1, 2, . . . , N.
Then the following theorem shows how the solution to one problem can be
used to obtain the solution to the other.
Table 2.2 Mirror relationship between the LMM rate allocation problem PR and the
LMM node lifetime problem PL .
PR PL
ri (optimization variable) ri = R (constant)

ti = T (constant) ti (optimization variable)
Total bit volume at AFN i: ri T = ti R
Theorem 2.1
Mirror relation For a given node lifetime requirement T for all nodes
under Problem PR and a given local bit rate R for all nodes under Problem
PL , we have the following relationship between the solutions to the LMM
rate allocation problem PR and the LMM node lifetime problem PL .
(i) Suppose that we have solved Problem PR and obtained the LMM-
optimal rate allocation ri for each node i (i = 1, 2, . . . , N ). Then under
PL , the LMM node lifetime ti for node i is given by
ri T
ti =. (2.14)
R
(ii) Suppose that we have solved Problem PL and obtained the LMM-
optimal node lifetime ti for each node i (i = 1, 2, . . . , N). Then under PR ,
the LMM rate allocation ri for node i is given by
ti R
ri = . (2.15)
T
Table 2.2 shows the mirror relationship between solutions to Problems PR

and PL .
Proof. We prove (i) and (ii) in Theorem 2.1 separately.

(i) We organize our proof into two parts. First, we show that the ti s are feasible
node lifetimes in terms of flow balance and energy constraints at each node i
(i = 1, 2, . . . , N). Then we show that this solution indeed achieves the LMM-
optimal node lifetime.
Feasibility Since we have obtained the solution to Problem PR , we have one
feasible flow routing solution for sending bit rates ri , i = 1, 2, . . . , N, to the
base station. Under Problem PR , the bit volumes (Vij s and ViB s) must meet
the following equalities under the LMM-optimal rate allocation:

ViB + Vij Vki = ri T ,
1j N,j =i 1kN,k=i

Vki + Cij Vij + CiB ViB = ei .
1kN,k=i 1j N,j =i
29 2.6 A mirror result
Now, replacing ri T by ti R, we see that the same bit volume solution under PR
yields a feasible bit volume solution to the node lifetime problem under PL .
Consequently, we can use Algorithm 2.1 to obtain the flow routing solution to
Problem PL under the bit volume solution to Problem PL and this verifies that
ti , i = 1, 2, . . . , N, is a feasible solution to Problem PL .
Optimality The proof is based on contradiction. But first, lets give some
notations. To prove that the ti -values obtained via (2).14 are indeed LMM-
optimal for Problem PL , we sort ri , i = 1, 2, . . . , N , under Problem PR in
nondecreasing order and denote it as [g1 , g2 , . . . , gN ]. We also introduce
a node index I = [i1 , i2 , . . . , iN ] for [g1 , g2 , . . . , gN ]. For example, i3 = 7
means that g3 actually corresponds to the rate of AFN 7, i.e., g3 = r7 .
Since ti is proportional to ri through the relationship (ti = R T
ri ), listing
ti , i = 1, 2, . . . , N, according to I = [i1 , i2 , . . . , iN ] will yield a sorted (in
nondecreasing order) lifetime list, denoted as [1 , 2 , . . . , N ]. We now prove
that if [1 , 2 , . . . , N ] is not LMM-optimal for Problem PL , then we will have
that [g1 , g2 , . . . , gN ] is not LMM-optimal, which is a contradiction.
Suppose that [1 , 2 , . . . , N ] is not LMM-optimal for Problem PL . Assume
that the LMM-optimal lifetime vector to Problem PL is [1 , 2 , . . . , N ]
(sorted in non-decreasing order) with the corresponding node index being
I = [i1 , i2 , . . . , iN ]. Then, by Definition 2.2, there exists a k such that j = j
for 1 j k 1 and k > k .
We now claim that if ti , i = 1, 2, . . . , N, is a feasible solution to Problem

PL , then ri obtained via ri = tiTR , i = 1, 2, . . . , N, is also a feasible solution
to Problem PR . The proof to this claim follows identically as above. Using
this result, we can obtain a corresponding feasible solution [g1 , g2 , . . . , gN ]
with gi = Ti R and the node index I for Problem PR . Hence, we have
R R
gj = jT = jT = gj for 1 j k 1 but gk = kTR > kTR = gk . That is,
[g1 , g2 , . . . , gN ] is not LMM-optimal and this leads to a contradiction.
(ii) The proof for this part is similar to that for (i) and is left as a homework
exercise.
This mirror relationship offers important insights on system performance

issues, in addition to providing solutions to the LMM rate allocation and the
LMM node lifetime problems. For example, in Section 2.2, we pointed out the
potential bias (fairness) issue associated with the network capacity maximiza-
tion objective (i.e., sum of rates from all nodes). It is interesting to see that
a similar fairness issue exists in the node lifetime problem. In particular, the
objective of maximizing the sum of node lifetimes among all nodes also leads
to a bias (or fairness) problem because this objective would only favor those
nodes that consume energy at a small rate. As a result, certain nodes will have
much larger lifetimes, while some other nodes will be penalized with much
smaller lifetimes, although the sum of node lifetimes is maximized.
Y (m) Y (m)
500 500
X (m) X (m)
500 0 500 500 0 500
500 500
(a) A 10-AFN network. (b) A 20-AFN network.
Figure 2.3 Network topologies

Table 2.3 Node coordinates for a 10-AFN network.
for the 10-AFN and 20-AFN
networks.
AFN i (xi , yi ) (in meters) AFN i (xi , yi ) (in meters)
1 (400, 320) 6 (500, 100)

2 (300, 440) 7 (400, 0)
3 (300, 420) 8 (420, 120)
4 (320, 100) 9 (200, 140)
5 (120, 340) 10 (220, 340)
2.7 Numerical results
In this section, we use numerical results to illustrate our SLP-PA algorithm

to the LMM rate allocation problem and compare it with other approaches.
We also use numerical results to illustrate the mirror relationship between the
LMM rate allocation problem and the LMM node lifetime problem.
We consider two network topologies, one with 10 AFNs and the other with
20 AFNs. Under both topologies, the base station B is located at the origin,
while the locations for the 10 or 20 AFNs are randomly generated over a
1000m 1000m square area. See Figs. 2.3(a) and (b) and Tables 2.3 and 2.4
respectively for the two network topologies. We set 1 = 50 nJ/b, 2 = 0.0013
pJ/b/m4 [67], and = 4 [129] in this study.
2.7.1 SLP-PA algorithm to the LMM rate allocation problem

We will compare SLP-PA with the naive approach SLP-ER (see Section 2.3.2)
that uses serial LP to blindly solve the LMM rate allocation problem and
31 2.7 Numerical results
Table 2.4 Node coordinates for a 20-AFN network.
AFN i (xi , yi ) (in meters) AFN i (xi , yi ) (in meters)
1 (200, 130) 11 (110, 230)

2 (400, 430) 12 (210, 0)
3 (100, 420) 13 (210, 320)
4 (0, 430) 14 (300, 480)
5 (410, 440) 15 (420, 470)
6 (200, 230) 16 (120, 240)
7 (400, 490) 17 (220, 440)
8 (410, 300) 18 (220, 240)
9 (100, 310) 19 (500, 110)
10 (10, 140) 20 (20, 330)
Table 2.5 Rate allocation under three approaches for the 10-AFN network.
i SLP-PA SLP-ER MaxCap

(sorted node gi AFN gi AFN gi AFN
index) (Kb/s) (Kb/s) (Kb/s)
1 0.1023 3 0.1023 1 0.0553 2

2 0.1023 6 0.1023 2 0.0627 3
3 0.1023 7 0.1023 3 0.0646 1
4 0.1536 5 0.1023 6 0.0658 6
5 0.2941 1 0.1023 7 0.1222 8
6 0.2941 2 0.1536 5 0.1653 10
7 0.2941 4 0.1536 8 0.1736 7
8 0.2941 8 0.1536 10 0.2628 5
9 0.2941 9 0.6563 4 0.3513 4
10 0.2941 10 0.6563 9 1.2398 9
performs energy reservation during each iteration. As discussed in Sec-

tion 2.3.2, the SLP-ER approach will not give the correct final solution to the
LMM rate allocation problem.
We will also compare our SLP-PA algorithm to the Maximum-Capacity
(MaxCap) approach (see Section 2.3.2). As discussed in the beginning of
Section 2.3.2, the rate allocation under the MaxCap approach can be extremely
biased and in favor of only those AFNs that consume the least power along
their data paths toward the base station.
10-AFN network We assume that the initial energy at each AFN is 50 KJ
and that the network lifetime requirement is 100 days under the LMM rate
allocation problem. The power consumption is for transmission and reception
defined in (2.2) and (2.4), respectively.
Table 2.5 shows the rate allocation for the AFNs under each approach,
which is also plotted in Fig. 2.4(a). The sorted node index corresponds to
the sorted rates among the AFNs in nondecreasing order. Clearly, among the
3.5
1.2
3 SLPPA SLPPA
SLPER 1 SLPER
2.5 MaxCap MaxCap
Rate (kb/s)
0.8
Rate (kb/s) 2
0.6
1.5
1 0.4
0.5 0.2
0 0
2 4 6 8 10 12 14 16 18 20 1 2 3 4 5 6 7 8 9 10
Sorted node index Sorted node index
(a) A 10-AFN network. (b) A 20-AFN network.
Figure 2.4 Rate allocation

under the SLP-PA, SLP-ER, and
three rate allocation approaches, only the rate allocation under SLP-PA meets
MaxCap approaches for a the LMM-optimal rate allocation definition (see Definition 2.1). Specifically,
10-AFN network and a 20-AFN comparing SLP-PA with SLP-ER, we have g1SLP-PA = g1SLP-ER , g2SLP-PA =
network. g2SLP-ER , g3SLP-PA = g3SLP-ER , and g4SLP-PA > g4SLP-ER ; comparing SLP-PA with
MaxCap
MaxCap, we have g1SLP-PA > g1 .
We also observe, as expected, a severe bias in the rate allocation under
MaxCap
the MaxCap approach. In particular, g10 alone accounts for over 48%
of the sum of total rates among all the AFNs. For the three approaches, we
MaxCap MaxCap
have g1SLP-PA = g1SLP-ER > g1 SLP-PA
and g10 < g10SLP-ER
< g10 . In other
words, the rate allocation vector under the SLP-PA algorithm has the small-
est rate difference between the smallest rate (g1 ) and the largest rate (g10 ), i.e.,
g10 g1 , among the three approaches. In addition, although g1SLP-PA = g1SLP-ER
for the first level rate allocation, the minimum node set for g1SLP-PA is smaller
than the minimum node set for g1SLP-ER , i.e., |S1SLP-PA | = 3 < |S1SLP-ER | = 5.
This confirms that the naive SLP-ER approach does not necessarily produce
the correct solution to the LMM rate allocation problem.
20-AFN network For the 20-AFN network (Table 2.4), we assume that the
initial energy at each AFN is 50 KJ and that the network lifetime require-
ment under the LMM rate allocation problem is 100 days. Table 2.6 shows the
sorted rate allocation under the three approaches, which are also displayed in
Fig. 2.4(b). It can be easily verified that all the observations for the 10-AFN
network also hold here.
2.7.2 Mirror results

We now use numerical results to verify the mirror relationship between the
LMM rate allocation problem (PR ) and the LMM node lifetime problem (PL )
Table 2.6 Rate allocation under three approaches for the 20-AFN network.
i SLP-PA SLP-ER MaxCap

(Sorted node gi AFN gi AFN gi AFN
index) (Kb/s) (Kb/s) (Kb/s)
1 0.3182 2 0.3182 2 0.0278 7

2 0.3182 7 0.3182 3 0.0282 15
3 0.3182 8 0.3182 4 0.0340 5
4 0.3182 11 0.3182 5 0.0374 2
5 0.3182 12 0.3182 6 0.0433 14
6 0.3182 14 0.3182 7 0.0648 19
7 0.3182 15 0.3182 8 0.0668 8
8 0.3182 16 0.3182 11 0.0760 17
9 0.3182 17 0.3182 12 0.1280 3
10 0.3182 18 0.3182 14 0.1301 4
11 0.3182 19 0.3182 15 0.2070 13
12 0.5694 5 0.3182 16 0.3714 20
13 1.3099 1 0.3182 17 0.3941 9
14 1.3099 3 0.3182 18 0.3948 18
15 1.3099 4 0.3182 19 0.5135 6
16 1.3099 6 2.0344 1 0.8524 16
17 1.3099 9 2.0344 9 1.0441 11
18 1.3099 10 2.0344 10 1.3588 1
19 1.3099 13 2.0344 13 2.2446 12
20 1.3099 20 2.0344 20 10.4362 10
(see Section 2.6). Again, we use the 10-AFN and 20-AFN network configura-
tions in Figs. 2.3(a) and (b), respectively. For both networks, we assume that
the initial energy at each AFN is 50 KJ and that the network lifetime require-
ment under the LMM rate allocation problem is T = 100 days. Under PL , we
assume that the local bit rate for all AFNs is R = 0.2 Kb/s.
To verify the mirror relationship (Theorem 2.1), we perform the following
calculations. First, we solve the LMM rate allocation problem (PR ) and the
LMM node lifetime problem (PL ) independently with the above initial condi-
tions using the SLP-PA algorithm. Consequently, we obtain the LMM-optimal
rate allocation (ri for each AFN i) under PR and the LMM-optimal node life-
time (ti for each AFN i) under PL . Then we compute T ri and R ti sepa-
rately for each AFN i and examine if they are equal to each other.
The results for the LMM-optimal rate allocation (ri , i = 1, 2, . . . , 10) and
the LMM-optimal node lifetime (ti , i = 1, 2, . . . , 10) for the 10-AFN network
are shown in Table 2.7. We find that T ri and R ti are exactly equal for all
AFNs, precisely as we would expect under Theorem 2.1. Similarly, the results
for the 20-AFN network are shown in Table 2.8.
Table 2.7 Mirror relationship T ri D R ti between the LMM rate allocation prob-
lem (PR ) and the LMM node lifetime problem (PL ) for the 10-AFN network.
AFN PR (T = 100 days) PL (R = 0.2 Kb/s)

i ri T ri ti R ti
1 0.2941 29.41 147.07 29.41

2 0.2941 29.41 147.07 29.41
3 0.1023 10.23 51.17 10.23
4 0.2941 29.41 147.07 29.41
5 0.1536 15.36 76.79 15.36
6 0.1023 10.23 51.17 10.23
7 0.1023 10.23 51.17 10.23
8 0.2941 29.41 147.07 29.41
9 0.2941 29.41 147.07 29.41
10 0.2941 29.41 147.07 29.41
Table 2.8 Mirror relationship T ri D R ti between the LMM rate allocation prob-
lem (PR ) and the LMM node lifetime problem (PL ) for the 20-AFN network.
AFN PR (T = 100 days) PL (R = 0.2 Kb/s)

i ri T ri ti R ti
1 1.3099 130.99 654.94 130.99

2 0.3182 31.82 159.10 31.82
3 1.3099 130.99 654.94 130.99
4 1.3099 130.99 654.94 130.99
5 0.5694 56.94 284.71 56.94
6 1.3099 130.99 654.94 130.99
7 0.3182 31.82 159.10 31.82
8 0.3182 31.82 159.10 31.82
9 1.3099 130.99 654.94 130.99
10 1.3099 130.99 654.94 130.99
11 0.3182 31.82 159.10 31.82
12 0.3182 31.82 159.10 31.82
13 1.3099 130.99 654.94 130.99
14 0.3182 31.82 159.10 31.82
15 0.3182 31.82 159.10 31.82
16 0.3182 31.82 159.10 31.82
17 0.3182 31.82 159.10 31.82
18 0.3182 31.82 159.10 31.82
19 0.3182 31.82 159.10 31.82
20 1.3099 130.99 654.94 130.99
2.8 Chapter summary
In this chapter, we reviewed LP and showed how it can be employed to solve

certain problems in wireless networks. Although the LP methodology itself
is rather basic and straightforward, special care is still needed to ensure that
35 2.8 Chapter summary
it is used correctly, as we demonstrated this in the case study in this chapter.

We introduced the parametric analysis (PA) technique, which is very useful
in its own right. We refer readers to [10] for other advanced techniques in
LP. We also introduced the concept of lexicographical max-min (LMM) rate
allocation, which can be employed as a fairness criterion for other problems in
wireless networking research. Our case study is rather interesting as it showed
that even in LP-based problem formulations, deep insights could be gained
once we dig deep into it. We hope the readers can exploit the LP method to its
fullest extent in their research endeavor, as we have demonstrated in this case
study.
Appendix: Optimal flow routing solution for LMM-optimal node lifetime
It is straightforward to develop an example similar to the one given in Sec-

tion 2.4.3 that demonstrates the nonuniqueness of the flow routing schedule.
Given that the optimal flow routing solution is nonunique, there are potentially
many flow routing solutions that can achieve the LMM-optimal lifetime vector.
In this appendix, we show a simple polynomial-time algorithm that provides
an LMM-optimal flow routing solution.
The main task in this algorithm is to define flows from the bit volumes (Vij
and ViB values), which are obtained upon the completion of the last iteration
in the LMM rate problem with our SLP-PA algorithm. Note that the bit vol-
umes obtained here represent the total amount of bit volume being transported
between the nodes during [0, n ], where n = N is the time when the last
group of nodes drains up their energy. The main result here is that if we let the
total amount of out-going flow at a node be distributed proportionally to the bit
volumes on each out-going link for all the remaining alive nodes at each stage,
then we can achieve the drop points 1 , 2 , . . . , n as well as the correspond-
ing minimum node sets S1 , S2 , . . . , Sn . The algorithm is formally described
in Algorithm 2.1 and proof of its correctness can be found in [69].
As shown in this algorithm, for each time interval (l1 , l ], l =
1, 2, . . . , n, we initialize Ul as the set of remaining alive nodes at this stage,

which is represented by Ul = S l1 h=0 Sh . For these nodes, we compute flow
routings by starting with the boundary nodes and then move to the interior
nodes. More precisely, we calculate the flow routing for a node i if and only if
we have calculated the flow routing for each node m that has traffic going into
node i. The out-going flow from node i is calculated by distributing the aggre-
gated flow proportionally according to the overall bit volume along its out-
going radio links. As an example, suppose that during (4 , 5 ], node 2 receives
an aggregated flow with rate 2 Kb/s and generates 0.4 Kb/s locally. Assume
that V24 = 100 Kb, V25 = 200 Kb, and V2B = 300 Kb over [0, n ]. Then the
Algorithm 2.1 An optimal flow routing solution

Upon the completion of the SLP-PA algorithm for the LMM-optimal lifetime
vector, we have the drop points (in strictly increasing order) 1 , 2 , . . . , n ,
the corresponding minimum node sets S1 , S2 , . . . , Sn , and the total amount
of bit volume on each radio link (i.e., Vij s and ViB s). The following algo-
rithm gives an LMM-optimal flow routing solution for time interval (l1 , l ],
where 0 = 0 and l = 1, 2, . . . , n:

1. Denote Ul = S l1 h=0 Sh , with S0 = . Initialize all flows to zero, i.e.,
(l) (l)
fij = 0, fiB = 0 for 1 i, j N, j = i.
2. If Ul = , then stop, else choose a node i from Ul such that
node i does not receive data from any other node, or
all nodes from which node i receives data are not in Ul .
3. The flow routing at node i during (l1 , l ] is then defined as
Vij
(l)
fij = fki(l) + ri (1 j N, j = i)
ViB + m=i Vim
k=i
ViB
(l)
fiB = fki(l) + ri ,
ViB + m=i Vim
k=i
where the fki(l) -values, if not zero, have all been defined before calculating
the flow routing for node i.
4. Let Ul = Ul {i} and go to Step 2.
(5) (5)
out-going flow at node 2 is routed as follows: f24 = 0.4 Kb/s, f25 = 0.8
(5)
Kb/s, and f2B = 1.2 Kb/s.
2.9 Problems
2.1 In an LP problem, suppose that the number of variables and of con-

straints are Nv and Nc respectively in the standard form. What is the complex-
ity of the solution to this problem?
2.2 For flow routing in a sensor network, we allow an outgoing flow from a
node to be split into subflows, each routed to a different next hop node (e.g.,
(2).5). Please explain the benefits of flow splitting (1) in terms of traffic flow
in the network, and (2) in terms of complexity in mathematical formulation
(when compared to single-path routing without flow splitting).
2.3 Why does the MaxCap LP formulation lead to unfairness in rate assign-
ment in ri , 1 i N ? Give an example of a small network and construct a
37 2.9 Problems
numerical result to illustrate this point. You may use numbers with normalized
units (dimensionless) in this numerical example.
2.4 Suppose that we obtain the following three sorted rate vectors corre-
sponding to three feasible rate allocations (each following a different solution
approach):
a = [0.10, 0.10, 0.10, 0.15, 0.29, 0.29, 0.29, 0.29, 0.29, 0.29]
b = [0.10, 0.10, 0.10, 0.10, 0.10, 0.15, 0.15, 0.15, 0.65, 0.65]
c = [0.05, 0.06, 0.06, 0.06, 0.12, 0.16, 0.17, 0.26, 0.35, 1.23]
Which two vectors are definitely not LMM-optimal? Why?
2.5 Discuss the similarity and the difference between the max-min rate allo-
cation problem and the LMM rate allocation problem.
2.6 In this chapter, two incorrect solution approaches for the LMM rate allo-
cation are presented. Explain why these two approaches cannot provide an
LMM-optimal rate allocation.
2.7 What is the objective of using parametric analysis (PA) in the solution of
the LMM rate allocation problem? Explain the basic idea of PA.
2.8 Is the minimum node set for each rate level under the LMM-optimal rate
allocation unique? Is the flow routing solution corresponding to the LMM-
optimal rate allocation unique?
2.9 What is the connection between the LMM rate allocation problem and
the LMM node lifetime problem in terms of mathematical formulation? What
is the mirror relationship between the two problems and when does it occur?
What is the application of this mirror relationship?
2.10 For Theorem 2.1, give a proof for the second statement:
Suppose that we have solved Problem PL and obtained the LMM-optimal
node lifetime ti for each node i (i = 1, 2, . . . , N). Then under PR , the LMM
rate allocation ri for node i is
ti R
ri = .
T
CHAPTER
3 Convex programming and

applications
They can because they think they can.
Virgil
3.1 Review of key results in convex optimization
In the previous chapter, we described LP and illustrated its application to solve

some interesting problems in wireless networks. In this chapter, we describe
convex programming [11], which is a popular tool to solve a wide range of
problems in wireless networks. In terms of problem space, LP can be viewed
as a special case of convex optimization. To facilitate our description, we define
the following terms:
1. Convex set: A set is convex if for any two of its elements z1 and z2 and for
any [0, 1], z1 + (1 )z2 is also an element of this set. For example,
the set {(x, y) : x 2 + y 2 1} is a convex set but the set {(x, y) : 1 x 2 +
y 2 2} is not a convex set.
2. Convex and concave functions: A function f (x) is a convex function
if for any x1 and x2 and any [0, 1], f (x1 + (1 )x2 ) f (x1 ) +
(1 )f (x2 ), where x can be a single variable or a vector of variables. A
function f (x) is a concave function if for any x1 and x2 and any [0, 1],
f (x1 + (1 )x2 ) f (x1 ) + (1 )f (x2 ). For example, f (x) = x 2 is
a convex function, f (x) = ln x is a concave function, and f (x) = x 3 is
neither convex nor concave. Observe that if f (x) is a convex function, then
f (x) is a concave function, and vice versa.
3. Affine function: A function f (x) is an affine function if it is both convex
and concave. It can be verified that an affine function is a linear function
plus some real constant, i.e., f (x) = ax + b for some constants a and b.
38
39 3.1 Review of key results in convex optimization
A convex programming problem usually has the following form:
CP Maximize f (x)
subject to g(x) 0
h(x) = 0
x X,
where x is a vector of continuous variables, f (x) is a concave objective func-

tion, g(x) is a vector of convex functions, h(x) is a vector of affine functions,
and X is a convex set. Note that a maximization problem can be readily con-
verted into a minimization problem (and vice verse) by multiplying the coeffi-
cients of the objective function by 1, i.e.,
Maximum f (x) = Minimum f (x),
where f (x) is a convex function since f (x) is a concave function.

It is easy to verify that any LP is a convex program but a mixed-integer linear
programming (MILP) is not. In general, any optimization problem with inte-
ger/binary variables is not a convex program because of the discrete nonconvex
nature of the set of integer/binary variable values.
A desirable property shared by convex programming problems is that a local
optimal solution is also a global optimal solution [11]. Therefore, for a given
convex programming problem, we can apply any efficient local optimization
approach, e.g., deflected or conjugate (sub)gradient methods or gradient pro-
jection methods [11], to improve the current solution iteratively and eventually
converge to an optimal solution. In particular, interior point methods can obtain
a (1 )-optimal solution in O(log10 ( 1 )) time [114], where is the gap to the
optimal solution. Due to this low-order complexity, we can obtain a solution
very close to the optimal solution with a low complexity, e.g., when = 109 ,
the complexity is only O(9). Thus, in practice, we can regard a convex pro-
gramming problem as a problem that can be solved optimally with a low-order
complexity.
For the foregoing convex programming CP, its Lagrangian dual problem is
given by
LD Minimize (u, v)
subject to u 0,
where u is the Lagrangian multiplier (or dual variable) associated with the
constraint g(x) 0, the unconstrained variable v is the Lagrangian multiplier
associated with the constraint h(x) = 0, and where
(u, v) max{f (x) uT g(x) vT h(x)}.

xX
40 Convex programming and applications
The function f (x) uT g(x) vT h(x) is called the Lagrangian function and
provides an upper bound on f (x) for any feasible solution to CP. Hence,
(u, v) max {f (x) uT g(x) vT h(x)}

xX,g(x)0,h(x)=0
max {f (x)}
xX,g(x)0,h(x)=0
for any u 0, and therefore v(LD) v(CP ), where v(p) denotes the optimal
value for any problem p. This latter relationship is known as the weak dual-
ity property. Under certain constraint qualifications (e.g., Slaters condition,
which requires the existence of an x X such that h(x) = 0, g(x) < 0, and the
origin belongs to the interior of the set {h(x) : x X}, convex programming
problems have the strong duality relationship v(LD) = v(CP ), i.e., both the
original primal problem and its dual problem have the same optimal objective
value [11]. Moreover, given an optimal solution to one problem, the optimal
solution to the other can be constructed using a suitable algorithmic proce-
dure [11]. Therefore, a primal problem can be possibly solved in its dual
domain.
An excellent reference on convex optimization is [11]. In the rest of this
chapter, we offer a case study to show how convex optimization can be applied
to solve a problem for a multi-hop MIMO network.
3.2 Case study: Cross-layer optimization for multi-hop

MIMO networks
As a case study for convex optimization, we consider a cross-layer opti-
mization problem for multi-hop MIMO networks. Under traditional single
transmit/receive antenna node architecture (also known as single-input single-
output or SISO), channel capacity is limited by the well-known Shannon
capacity: C = log(1 + SNR) per Hz, where SNR represents the signal-to-noise
ratio. However, thanks to the work by Winters [169], Foschini and Gans [41],
and Teletar [155] on multiple-input multiple-output (MIMO), much higher
spectrum efficiency and capacity gain can now be achieved. The benefits of
substantial improvements in capacity at no additional cost of spectrum have
already positioned MIMO as one of the major breakthroughs in modern wire-
less communications [17; 48; 122]. To date, MIMO has found its way into
many wireless network standards, such as wireless LAN (802.11n), WiMAX
access networks (802.16e), and 4G systems.
Employing MIMO in a multi-hop network is not trivial. As discussed by
Winters [170], a multi-hop MIMO network with each node equipped with
M antennas does not necessarily mean that the network throughput is also
increased by M-fold. The potential network throughput gain with the use of
MIMO depends on the coordinated mechanisms at the physical, link, and
41 3.3 Network model
network layers. As a result, joint optimization across multiple layers is not

only desirable, but also necessary.
Consider a set of user communication sessions in a multi-hop MIMO net-
work. The objective is to maximize some network utility function, which is
a function of active sessions throughput. A number of issues will be con-
sidered in this problem, including (i) at the physical layer, how to determine
the optimal input covariance matrix at each transmitting node, (ii) at the link
layer, how to optimally allocate bandwidth on different outgoing links, (iii)
at the network layer, how to perform optimal flow routing in the network, and
(iv) at the transport layer, how to optimally establish a session between a source
and destination nodes.
We will show that this challenging cross-layer optimization problem can be
formulated into a convex programming problem. Just like other convex pro-
gramming problems, this problem can be solved in its dual domain. But for
this particular problem, we will show that it has some special structure that
allows us to decompose the problem in the dual domain into two subproblems.
That is, the dual problem can be de-coupled in to a transport-network layer
subproblem and a link-physical layer subproblem. Both subproblems are also
convex programs. Then, we propose a cutting-plane method for the dual prob-
lem. We show that our method allows an easy recovery of a primal feasible and
optimal solution.
We organize the rest of the sections as follows. In Section 3.3, we describe
the network model and problem formulation. Section 3.4 presents a decom-
position framework and resulting subproblems. In Section 3.5, we show the
cutting-plane method to solve the dual problem. Section 3.6 describes a method
to recover an optimal primal solution. Section 3.7 presents numerical results.
3.3 Network model
We assume orthogonal channels on all links in the network (similar to that in

[27; 83; 105]). Note that orthogonal channels do not require as many channels
as the number of active links in the network since we can reuse channels on
links that are spatially far away from each other. This is called spatial reuse
and is commonly used in wireless networks to improve channel efficiency.
Since efficient channel assignment algorithms have been well studied in the
literature, we will assume that channel assignment has been taken care of and
thus will not be part of our problem formulation.
Under orthogonal channels, each link in the network is assigned some chan-
nel. A node may transmit to different receiving nodes on different links, each
with a non-overlapping channel. Further, a node can simultaneously transmit
and receive on different channels and does not cause any self-interference.
In this chapter, we focus on how to jointly optimize input covariance matri-

ces determination at the physical layer, bandwidth allocation at the link layer,
sessions flow rate assignment at the transport layer and flow routing at the
network layer.
We first introduce the notation for matrices, vectors, and complex scalars.
We use boldface to denote matrices and vectors. For a matrix A, A denotes
the conjugate transpose. Tr{A} denotes the trace of A. Diag{A1 , . . . , An } repre-
sents the block diagonal matrix with matrices A1 , . . . , An on its main diagonal.
We denote I as the identity matrix with dimension determined from the context.
A
0 represents that A is Hermitian and positive semidefinite (PSD). 1 and 0
denote vectors whose elements are all ones and zeros, respectively, where their
dimensions are determined from the context. (v)i represents the ith entry of a
vector v. For a real vector v and a real matrix A, v 0 and A 0 mean that all
entries in v and A are nonnegative, respectively. We let ei be the unit column
vector where the ith entry is 1 and all the other entries are 0, and where the
dimension of ei is determined from the context. The operator , represents
the inner product of two vectors or matrices. Table 3.1 lists the notations used
in this chapter.
3.3.1 MIMO input covariance matrices

Denote xli as the signal sent through the ith antenna at the transmitter of link l.
Let matrix Ql represent the input covariance matrix of link l, where the (i, j )th
j
element is E[xli (xl ) ]. A covariance matrix is Hermitian and PSD. Physically,
the diagonal elements of Ql represent power allocation at different antennas at
the transmitting node of link l. Then we have that

Tr{Ql } Pmax (1 n N ),
lO(n)
where O (n) is the set of outgoing links that from node n and
Pmax is the max-

imum transmit power of a node. We use the matrix Q Q1 Q2 . . . QL

CM(ML) to denote the collection of all input covariance matrices in the net-
work, where L is the number of links.
3.3.2 Link capacity and bandwidth allocation

Let the matrix Hl CMM represent the wireless channel gain matrix from
the transmitting node to the receiving node of link l, where M is the number
of antenna elements at each node. Suppose that Hl is known at the transmit-
ting node of link l. Although wireless channels in reality are time varying, we
consider a constant channel model in this chapter, i.e., Hl s coherence time
is larger than the considered transmission period. This simplification is still of
much interest as it allows us to find the ergodic capacity for block-wise fading
channels [55].
Table 3.1 Notation.
Symbol Definition
1, 0 Vectors whose elements are all ones or zeros

A The node-link incidence matrix
BO(n) The bandwidth allocated at node n for transmission
BI(n) The bandwidth allocated at node n for reception
c(Wl , Ql ) The link capacity of a MIMO link l with bandwidth Wl
and input covariance matrix Ql
d(f ) The destination node of session f
ei Vector where the ith entry is 1 and all the other entries are 0
F The number of sessions in the network
Hl The channel gain matrix from the transmitting node to the
receiving node of link l
I The identity matrix with dimension determined from the
context
I (n) The set of incoming links to node n
L The set of all MIMO links in the network
L The number of links in the network
M The number of antenna elements at each node
N The set of nodes in the network
N The number of nodes in the network
O (n) The set of outgoing links from node n
Pmax The maximum transmit power of a node
Ql The input covariance
matrix of link l
Q = Q1 Q2 . . . QL , the collection of all input covariance
matrices in the network
rf The flow rate of session f
Rf The flow rate vector
of session f
R = R1 R2 . . . RF , the collection of all source-sink vectors
s(f ) The source node of session f
ul Lagrangian
multiplier

for link ls capacity constraint (3.10)
u = u1 u2 . . . uL , the collection of all dual variables
Wl The communication bandwidth of link l

T
W = W1 W2 . . . WL , the collection of all bandwidth
variables
(f )
yl The amount of flow rate on link l that is attributed to session f
y(f ) = [ y (f ) y (f ) . . . yL
(f ) ]T , the flow rate vector of session f on
1 2
L links

Y = y(1) y(2) . . . y(F ) , the collection of all flow rate vectors

l The ratio of path-loss coefficient and noise power for link l
The link capacity of a MIMO link l in an AWGN channel can be com-

puted as

c(Wl , Ql ) Wl log2 det I + l Hl Ql Hl , (3.1)
where Wl represents the channel bandwidth of link l and l is the ratio of

path-loss coefficient and noise power for link l. It can be readily verified that
c(Wl , Ql ) is a monotone increasing concave function for Wl > 0 and Ql
0.
We assume the sum of the bandwidth of all outgoing links at a node n cannot
exceed BO(n) , the bandwidth at this node that is allocated for transmission. We
have

Wl BO(n) (1 n N ).
lO(n)
Similarly, we assume the sum of the bandwidth of all incoming links at node n
cannot exceed BI(n) , the bandwidth at this node that is allocated for reception.
We have

Wl BI(n) (1 n N ).
lI(n)

T
We denote the vector W = W1 W2 . . . WL RL as the collection of all
bandwidth variables.
It is easy to see that the values of Wl and Ql , i.e., the allocation of bandwidth
and power on link l, directly affect the capacity of link l. The determination of
these values will be part of our cross-layer optimization problem.
3.3.3 Flow routing

The topology of a multi-hop MIMO network can be represented by a directed
graph, denoted by G = {N , L}, where N and L are the set of nodes and all pos-
sible MIMO links in the network, respectively. We assume that a link between
two nodes exists if the distance between the two is no greater than the trans-
mission range RT , i.e., L = {(n, m) : Dnm RT , n, m N , n = m}, where
Dnm represents the distance between nodes n and m and RT is determined by
a nodes maximum transmission power. We assume G is connected.
The network topology of G can be mathematically modeled as a node-link
incidence matrix (NLIM) A RNL [10], with entry anl being defined as:

1 if n is the transmitting node of link l,
anl = 1 if n is the receiving node of link l,

0 otherwise.
We assume there are F sessions in the network. We denote the source and
destination nodes of session f , 1 f F , as s(f ) and d(f ), respectively.
We allow flow splitting inside the network so as to offer more flexibility in
flow routing and better load balancing. Denote rf as the flow rate of session
f , 1 f F , which are optimization variables.
(f )
Denote yl 0 as the amount of flow on link l attributed to session f . For
each session f , we have the following flow balance at node n:

(f ) (f ) rf if n = s(f ),
yl yl = rf if n = d(f ), (3.2)

lO(n) lI(n) 0 otherwise,
where I (n) is the set of links that are incoming to node n.
To rewrite (3.2) compactly, we use Rf RN to denote the right-hand-side
(RHS) of constraints for a session f , i.e., (Rf )s(f ) = rf , (Rf )d(f ) =
rf , and (Rf )n = 0 for 1 n N, n = s(f ), n = d(f ). We define y(f ) =
[ y1(f ) y2(f ) . . . yL(f ) ]T RL as the flow vector on all L links attributed to ses-
sion f . With NLIM, for any session f , (3.2) can be written Ay(f ) = Rf .

as NF
In addition, denoting the matrix R R1 R2 . . . RF R as the col-
lection of all source-sink vectors Rf , we have Rf = Ref . The matrix R needs
to satisfy the following constraints:
(Ref )s(f ) = rf (1 f F )
1, Ref = 0 (1 f F )
Ref = s(f ),d(f ) 0 (1 f F ),
where the notation =x,y represents the component-wise equality of a vector

except at the xth and yth entries. Denote the matrix Y y(1) y(2) . . . y(F )
RLF as the collection of all flow vectors y(f ) . Then, the flow conservation law
for all sessions can be written as
AY = R.
Since all flows traversing a link cannot exceed this links capacity limit,
(f )
we have Ff =1 yl c(Wl , Ql ) for 1 l L. Using matrix-vector notation,
this can be compactly written as
1, YT el c(Wl , Ql ) (1 l L).
Note that link capacity c(Wl , Ql ) depends on bandwidth allocation and the
input covariance matrix. Also note that this constraint couples three layers (i.e.,
network, link, and physical).
3.3.4 Problem formulation

Our cross-layer optimization problem involves joint optimization of sessions
flow rate assignment at the transport layer, flow routing at the network layer,
bandwidth allocation at the link layer, and input covariance matrices at the
physical layer. For the objective function, we adopt the proportional fair-
ness utility function, ln(rf ), for each session f [79], which is to maximize
the sum of utilities of all sessions. Putting together the physical layer con-
straints in Section 3.3.1, the link layer constraints in Section 3.3.2, and the
transport-network layer constraints in Section 3.3.3, we have the following

problem formulation:

F
P : Maximize ln(rf )
f =1

subject to Tr{Ql } Pmax (1 n N ) (3.3)
lO(n)

Wl BO(n) (1 n N ) (3.4)
lO(n)

Wl BI(n) (1 n N ) (3.5)
lI(n)
Ref =s(f ),d(f ) 0 (1 f F ) (3.6)

1, Ref = 0 (1 f F ) (3.7)
(Ref )s(f ) = rf (1 f F ) (3.8)
AY = R (3.9)
1, Y el c(Wl , Ql )
T
(1 l L) (3.10)
rf , Y 0 (1 f F )
Ql
0, Wl 0 (1 l L),
where c(Wl , Ql ) is defined in (3.1).
It can be verified that Problem P is a convex programming by definition.
Such a convex problem can be solved by standard approaches in [11]. On the
other hand, Problem P can also be solved in its dual domain. In particular,
we will show that the dual problem can be decomposed into subproblems,
each of which can be solved separately. Further, such approach is shown to
be amenable to distributed implementation. Given the significance of such a
dual decomposition approach in the literature [79; 90; 98], we examine this
approach in detail in the rest of this chapter.
3.4 Dual problem decomposition
In general, it is always better to decompose a large optimization problem into

smaller subproblems and solve each subproblem separately in a suitable coor-
dinated fashion. We find that in Problem P, the network layer variables and
the link/physical layer variables are coupled only through the link capacity
constraints (3.10). Thus, we may decompose the problem into one network
layer problem and one link-physical layer problem. But Problem P itself can-
not be decomposed. However, as we discussed in Section 3.1, we can solve
Problem P in its dual domain. In this section, we show how to decompose the
Lagrangian dual formulation of Problem P and solve the smaller subproblems.
47 3.4 Dual problem decomposition
Generally, given a nonlinear programming problem, several different

Lagrangian dual problems can be constructed depending on which constraints
are chosen to be dualized, i.e., associated with the Lagrangian dual variable
[11]. If the strong duality result holds, then all of these dual problems have the
same optimal objective value as the original primal problem. For our cross-
layer optimization problem, there is a particular dual formulation that can be
decomposed. Further, an optimal solution to the original primal problem can
be constructed from an optimal solution to the dual problem.
We first map Problem P to the convex programming problem CP discussed
in Section 3.1 and then obtain its dual problem. Note that in Problem P,
the constraints (3.3) are power (physical layer) constraints; the constraints
(3.4) and (3.5) are bandwidth allocation (link layer) constraints; the con-
straints (3.6)(3.8) are flow routing (network layer) constraints; the con-
straint (3.9) is a coupling constraint that involved both the transport layer
and the network layer; and the constraints (3.10) are coupling constraints
that involve both the network layer and the physical/link layers. Thus, to
obtain a dual problem that can be decomposed into smaller subproblems,
we let g(x) = 1, YT el c(Wl , Ql ). Therefore, we have x = (R, Y, Q, W),

f (x) = Ff =1 ln(rf ), g(x) = 1, YT el c(Wl , Ql ), and

AY = R

Ref =s(f ),d(f ) 0 (1 f F )

1, Ref = 0 (1 f F )

(Re ) (1 f F )

f s(f ) = rf

X = (R, Y, Q, W) lO(n) Tr{Ql } Pmax (1 n N ) .

lO(n) Wl BO(n) (1 n N )

lI(n) Wl BI(n)
(1 n N )

rf , Y 0
(1 f F )

Q
0, W 0
l l (1 l L)
Note that there is no constraint h(x) = 0 in Problem P. Then its dual problem
is given by
Minimize (u)
subject to u 0,
where

F
L
(u) = max ln rf + ul c(Wl , Ql ) 1, YT el .
(R,Y,Q,W)X
f =1 l=1
(3.11)
It is easy to recognize that, for any given Lagrangian multiplier u, the
Lagrangian can be separated into two terms:
(u) = tsptnet (u) + linkphy (u),

where tsptnet and linkphy are two subproblems corresponding to the

transport-network layer and the link-physical layer respectively:

tsptnet (u) Maximize f ln rf l ul 1, YT el
subject to AY = R
Ref =s(f ),d(f ) 0 (1 f F )
1, Ref = 0 (1 f F )
(Ref )s(f ) = rf (1 f F )
rf , Y 0 (1 f F )

linkphy (u) Maximize l ul c(Wl , Ql )

subject to lO(n) Tr{Ql } Pmax (1 n N )

W BO(n) (1 n N )
lO(n) l
W
lI(n) l B I(n) (1 n N )
Ql
0, Wl 0 (1 l L).

For the transport-network layer subproblem, we can regard f ln rf as the

revenue (measured by a rate utility function) and l ul 1, YT el as the cost to
achieve session rates. For the link-physical layer subproblem, we can regard

l ul c(Wl , Ql ) as the revenue, which is measured by a sum of weighted link
capacities. Then, the Lagrangian dual problem of P can be transformed into
the following problem:
D: Minimize tsptnet (u) + linkphy (u)
subject to u 0.
The Lagrangian dual problem requires us to find the optimal u such that
tsptnet (u) and linkphy (u) can be minimized. Thus, we should be able
to evaluate the subproblem functions tsptnet (u) and linkphy (u) for any
given u. Note that to evaluate tsptnet (u) in the network layer subproblem,
the objective function is concave and all constraints are affine. Hence, this
subproblem is a convex program, as described in Section 3.1. Following the
same token, we can verify that the problem of evaluating linkphy (u) is also a
convex program. Therefore, both tsptnet (u) and linkphy (u) can be readily
solved by many efficient convex programming methods. In the next section,
we show how to solve Problem D.
3.5 Solving the Lagrangian dual problem
In this section, we show how to solve Problem D by a cutting-plane method.

The basic idea of this method is as follows. We first reformulate Problem D as
a linear program with an infinite number of constraints. Then we build a linear
relaxation for Problem D using a finite number of constraints. We begin with
a linear relaxation with only one constraint. During each iteration, we solve
49 3.5 Solving the Lagrangian dual problem
the relaxed problem, whose solution provides a lower bound for Problem D.
If this solution is not feasible to Problem D, we generate an additional linear
constraint to cut the current solution. With this additional linear constraint,
we obtain a tighter linear relaxation for the next iteration. Eventually, if we
obtain a linear relaxation using a finite number of constraints whose solution
to Problem D is feasible, then we would have derived an optimal solution to
Problem D.
We now present the details of the cutting-plane method. Let z = (u) =
F L
f =1 ln rf + l=1 ul c(Wl , Ql ) 1, Y el . Thus,
max(R,Y,Q,W)X T

for each u 0, the linear inequality z Ff =1 ln rf + L l=1 ul

c(Wl , Ql ) 1, Y el must hold for all (R, Y, Q, W) X. Problem
T
D is therefore equivalent to
Minimize z

subject to z f ln rf + l ul c(Wl , Ql ) 1, YT el , (3.12)
(R, Y, Q, W) Xu 0.
Note that each given point (R, Y, Q, W) yields a linear constraint for z and u.
Therefore, we can regard (3.12) as an LP problem having an infinite number of
constraints, with each constraint corresponding to a point (R, Y, Q, W) X.
Instead of enumerating all (infinite number of) points, we consider only a
finite number of points and add more points as needed. This gives us a relaxed
problem (since we consider only a subset of constraints). In the first iteration,
we begin with one point (R(0) , Y(0) , Q(0) , W(0) ) X.
In the kth iteration, we have k points (R(j ) , Y(j ) , Q(j ) , W(j ) ) X, j =
0, . . . , k 1, for which the relaxed problem is given as follows:
Minimize z
(j ) (j ) (j )
subject to z f ln rf + l ul c(Wl , Ql ) 1, Y(j )T el
(0 j < k) u 0.
(3.13)
This is an LP problem with k constraints and can be solved efficiently.

Denote (z(k) , u(k) ) as an optimal solution to this relaxed problem. Then z(k)
is a lower bound for the optimal z to the Lagrangian dual problem (3.12).
If this solution to problem (3.12) is feasible, then z(k) is an upper bound
for z . Since z(k) is both a lower and upper bound for z , we have have
z(k) = z , i.e., it is an optimal solution to problem (3.12). To check the fea-
sibility of (z(k) , u(k) ) to the Lagrangian dual problem (3.12), we must check
whether
(k)
z(k) max ln rf + ul c(Wl , Ql ) 1, YT el .
(R,Y,Q,W)X
f l
To do this, we consider the following problem:

(k)
Maximize f ln rf + l ul c(Wl , Ql ) 1, YT el
(3.14)
subject to (R, Y, Q, W) X.
Note that this problem is exactly the problem in (3.11). Therefore, as discussed
in Section 3.4, we can decompose this problem into two smaller subproblems
and solve each subproblem efficiently.
Algorithm 3.1 Cutting-plane algorithm for solving problem D

Initialization:
Find a point (R(0) , Y(0) , Q(0) , W(0) ) X.
Set k = 1.
Main Loop:
1. Solve the relaxed problem in (3.13) and obtain its solution (z(k) , u(k) ).
2. Solve (3.14) and obtain an optimal solution (R(k) , Y(k) , Q(k) , W(k) )
with objective value (u(k) ).
3. If z(k) (u(k) ), then stop and (z(k) , u(k) ) is an optimal dual solution.
4. Otherwise, add constraint (3.15) to the relaxed problem. Set k to k + 1,
and return to Step 1.
Denote (R(k) , Y(k) , Q(k) , W(k) ) as an optimal solution to (3.14) and (u(k) )
as the corresponding optimal objective value. If zk (u(k) ), then (z(k) , u(k) )
is feasible for (3.12) and is an optimal solution to this Lagrangian dual prob-
lem. Otherwise, for (z(k) , u(k) ), the inequality constraint in (3.12) is not satis-
fied for (R(k) , Y(k) , Q(k) , W(k) ). Thus, we add the constraint:
(k) (k) (k)

z ln rf + ul c(Wl , Ql ) 1, Y(k)T el (3.15)
f l
to (3.13), and solve the relaxed linear program again in the (k + 1)th itera-
tion (now with (k + 1) points). Obviously, (z(k) , u(k) ) will not show up in the
(k + 1)th iteration because it is cut off by (3.15). The cutting-plane algorithm
is summarized in Algorithm 3.1.
3.6 Constructing a primal optimal solution
Thus far, we have solved Problem D. We now show how we can construct an
optimal solution to Problem P.
Suppose that Algorithm 3.1 terminates at iteration k = K. Then we have
K + 1 points (R(j ) , Y(j ) , Q(j ) , W(j ) ) X, j = 0, . . . , K. The last solved LP
problem in Algorithm 3.1 is as follows:
Minimize z
(j ) (j ) (j )
subject to z f ln rf + l ul c(Wl , Ql ) 1, Y(j )T el
(0 j K) u 0.
(3.16)
As discussed in Section 2.1, when we solve (3.16), we have also solved its dual
problem, which is given as follows:
F (j )
Maximize K j =0 j f =1 ln(rf )
K
(j ) (j )
subject to j =0 j 1, Y(j )T el c(Wl , Ql ) 0, l
K (3.17)
j =0 j = 1
j 0, j.
By the LP duality theory, the optimal values j , for j = 0, . . . , K, are imme-
diately available after we solve (3.16). Furthermore, (3.16) and (3.17) have the
same optimal objective value z(K) .
To construct an optimal solution to Problem P, we first need to construct a
feasible solution. Such a feasible solution (R, Y, Q, W) should satisfy (3.10)
and (R, Y, Q, W) X. Since j , for j = 0, . . . , K, is a solution of (3.17), we

have K j =0 j (R , Y , Q , W ) X. Let
(j ) (j ) (j ) (j )

K
(R, Y, Q, W) = j (R(j ) , Y(j ) , Q(j ) , W(j ) ).
j =0
Then (R, Y, Q, W) X. If we can verify that (R, Y, Q, W) also satisfies

(3.10), then it is a feasible solution to Problem P. In fact, we can prove the fol-
lowing theorem for (R, Y, Q, W). The proof is given in the Appendix, which
is a simplified proof of Theorem 6.5.1 in [11] for our particular problem.
Theorem 3.1
(R, Y, Q, W) is a feasible and optimal solution to Problem P.
In this section, we present some numerical results to gain more quantitative

understanding of the problem and solution. We first describe the simulation
setting. As shown in Fig. 3.1, we consider 15 nodes uniformly deployed in an
1200m 1200m square area. Each node in the network is equipped with two
antennas. The maximum transmit power for each node n is set to Pmax = 20
dBm (0.1 W). The path-loss index is 3 and the noise power is 4 1021 W. The
total bandwidth at each node for transmission and reception are 20 MHz and
Figure 3.1 Network topology of 1200

a 15-node network.
N14
1000
800 N7 N5 N6
N11 N9
N10
(m)
600 N13 N15

N12
N3 N2
N8
400
N4
200 N1
0
0 200 400 600 800 1000 1200
(m)
30 MHz, respectively. There are three sessions in the network: node 14 to node
1, node 6 to node 10, and node 5 to node 4. The randomly generated channel
gain matrices are shown in Table 3.2.
After executing our solution procedure, we find that the optimal rates for
these three sessions are r1 = 125 Mb/s, r2 = 190.6 Mb/s, and r3 = 258 Mb/s.
The optimal objective value is 6.64. The routings and flow rates of sessions 1,
2, and 3 are shown in Figs. 3.2, 3.3, and 3.4, respectively. These figures show
that flow routings for sessions 1, 2, and 3 are all multi-path and multi-hop. It
can be easily verified that the flow rates in Figs. 3.2, 3.3, and 3.4 satisfy flow
conservation at each node.
Table 3.3 shows the optimal bandwidth allocation of the network. Table 3.4
shows the optimal input covariance matrices of the network. In Table 3.4, each
cell with four entries corresponds to a 2 2 Q-matrix, which represents an
input covariance matrix. Consider the transmission from N11 to N7 as an
example: Q(11,7) in Table 3.4 says that the power allocations to the two anten-
nas at N11 are 11.83 mW and 12.10 mW. Also, the signals sent through these
two antennas, denoted by x1 and x2 , should also be correlated with power
E[x1 x2 ] = (0.25 0.01i) mW and E[x2 x1 ] = (0.25 + 0.01i) mW.
It can be observed from Tables 3.3 and 3.4 that not every node allo-
cates its full power and total bandwidth for its outgoing links. For exam-
ple, N14 has only one outgoing link (14, 7). We can see that W(14,7) = 14.2
MHz and Tr(Q(14,7) ) = 79.47 mW, which shows N14 does not transmit at
full power and does not utilize all of its assigned bandwidth. Such a link
is not a bottleneck link. Even if the transmitter of a nonbottleneck link
increases its total transmit power and bandwidth, the end-to-end session rate
Table 3.2 Channel gain matrices.
0.79 0.62i 0.57 + 1.33i 0.47 2.12i 1.56 + 0.97i

H(1,4) H(8,13)
0.70 + 1.32i 0.17 0.57i 0.07 0.71i 0.42 + 0.18i
0.14 + 0.35i 0.17 0.19i 0.14 + 0.35i 1.75 + 0.81i
H(2,9) H(9,2)
1.75 + 0.81i 0.86 0.33i 0.17 0.19i 0.86 0.33i
0.21 + 0.67i 0.04 + 1.14i 0.20 0.15i 2.00 0.52i
H(2,12) H(9,6)
1.08 1.70i 0.03 + 0.68i 0.47 0.77i 0.46 + 0.71i
0.00 + 0.16i 1.12 + 0.36i 1.36 0.36i 0.43 0.45i
H(2,15) H(9,15)
0.79 0.85i 0.47 + 0.46i 0.37 0.37i 0.24 0.73i
0.73 + 0.58i 1.06 + 0.58i 0.27 0.03i 0.29 0.39i
H(3,4) H(10,7)
0.92 0.44i 0.07 1.40i 0.76 + 0.32i 0.92 1.02i
0.68 0.99i 0.66 0.08i 0.15 0.83i 0.76 0.59i
H(3,8) H(10,8)
0.20 + 0.99i 0.53 + 0.54i 0.00 0.47i 0.24 0.77i
1.52 0.51i 0.12 0.05i 0.31 0.72i 1.07 0.30i
H(3,11) H(10,13)
0.01 0.41i 2.17 0.35i 1.12 + 0.73i 0.92 0.47i
0.23 2.93i 0.29 + 0.70i 1.52 0.51i 0.01 0.41i
H(3,12) H(11,3)
0.94 0.39i 0.19 1.41i 0.12 0.05i 2.17 0.35i
0.46 + 0.63i 0.45 + 1.09i 1.02 + 1.43i 1.79 + 0.84i
H(3,13) H(11,5)
0.70 0.42i 0.11 0.23i 0.36 + 0.34i 1.02 0.32i
0.79 0.62i 0.70 + 1.32i 0.63 1.60i 0.75 0.28i
H(4,1) H(11,7)
0.57 + 1.33i 0.17 0.57i 0.25 1.17i 0.75 0.02i
0.73 + 0.58i 0.92 0.44i 1.19 0.47i 0.35 0.30i
H(4,3) H(11,12)
1.06 + 0.58i 0.07 1.40i 0.69 + 0.92i 0.54 + 0.26i
0.94 + 0.06i 0.57 0.75i 0.21 + 0.67i 1.08 1.70i
H(4,12) H(12,2)
0.01 + 0.65i 0.28 + 0.35i 0.04 + 1.14i 0.03 + 0.68i
0.26 1.07i 1.19 0.50i 0.23 2.93i 0.94 0.39i
H(5,7) H(12,3)
0.15 + 0.70i 1.25 0.45i 0.29 + 0.70i 0.19 1.41i
1.02 + 1.43i 0.36 + 0.34i 0.94 + 0.06i 0.01 + 0.65i
H(5,11) H(12,4)
1.79 + 0.84i 1.02 0.32i 0.57 0.75i 0.28 + 0.35i
0.20 0.15i 0.47 0.77i 1.19 0.47i 0.69 + 0.92i
H(6,9) H(12,11)
2.00 0.52i 0.46 + 0.71i 0.35 0.30i 0.54 + 0.26i
1.07 0.20i 1.12 0.38i 0.46 + 0.63i 0.70 0.42i
H(6,15) H(13,3)
0.37 0.19i 0.91 1.14i 0.45 + 1.09i 0.11 0.23i
0.26 1.07i 0.15 + 0.70i 0.95 + 1.00i 0.91 + 0.15i
H(7,5) H(13,7)
1.19 0.50i 1.25 0.45i 0.77 0.44i 1.64 0.18i
0.27 0.03i 0.76 + 0.32i 0.47 2.12i 0.07 0.71i
H(7,10) H(13,8)
0.29 0.39i 0.92 1.02i 1.56 + 0.97i 0.42 + 0.18i
0.63 1.60i 0.25 1.17i 0.31 0.72i 1.12 + 0.73i
H(7,11) H(13,10)
0.75 0.28i 0.75 0.02i 1.07 0.30i 0.92 0.47i
0.95 + 1.00i 0.77 0.44i 0.90 0.02i 0.56 0.22i
H(7,13) H(14,7)
0.91 + 0.15i 1.64 0.18i 0.96 + 0.18i 0.11 0.04i
0.90 0.02i 0.96 + 0.18i 0.00 + 0.16i 0.79 0.85i
H(7,14) H(15,2)
0.56 0.22i 0.11 0.04i 1.12 + 0.36i 0.47 + 0.46i
0.68 0.99i 0.20 + 0.99i 1.07 0.20i 0.37 0.19i
H(8,3) H(15,6)
0.66 0.08i 0.53 + 0.54i 1.12 0.38i 0.91 1.14i
0.15 0.83i 0.00 0.47i 1.36 0.36i 0.37 0.37i
H(8,10) H(15,9)
0.76 0.59i 0.24 0.77i 0.43 0.45i 0.24 0.73i
Figure 3.2 Routing and flow Session 1: N14 N1

rates of session 1 (in Mb/s). 1200
N14
1000 125
18.4 N5
N7
62.8 18.4
800 N6
N10 24.6
19.2 N11
15.2 N9
30.2 12.8
N13
43.4
(m)
600 47.6 9
1.4 N15
N12
38.6 N3 N2
N8 11.4
113.6
400
N4 125
200 N1
0
0 200 400 600 800 1000 1200
(m)

N14
1000
23.2 N5
N7 153.6
65.6 N6
800 23.2 N9
N10 99.2
51.4 N11 37
91.8
46.2 70.8 104.8 48.8
N13
17.8 N3
(m)
600 13.2 22.6 99.2 190.6 N15

N12 85.8
55.8 N2
N8 26
26
400
N4
200 N1
0
0 200 400 600 800 1000 1200
(m)
cannot be increased due to the minimum bottleneck link along the path.
In this network, it can be verified that N3, N7, N11, and N12 are bottle-
neck nodes. For example, at N11, W(11,3) + W(11,5) + W(11,7) + W(11,12) = 20
MHz and Tr(Q(11,3) ) + Tr(Q(11,5) ) + Tr(Q(11,7) ) + Tr(Q(11,12) ) = 100 mW.
This means that the power and bandwidth at N11 has been fully utilized and
cannot be further increased.

N14
1000
69.4 N5
N7
800 0.6 188.6 N6
N10 32.4
36.4
0.6
N11 N9
154.6 70.4
N13
33
600 N15
(m)
N12
N3 N2
N8 70.4
187.6
400
N4
200 N1
0
0 200 400 600 800 1000 1200
(m)
Table 3.3 Bandwidth allocation of the 15-node network (in MHz).
W(2,12) 8.20 W(6,9) 7.40 W(9,2) 7.00 W(12,4) 3.60

W(3,4) 3.40 W(6,15) 9.00 W(9,15) 6.80 W(12,11) 7.80
W(3,8) 2.80 W(7,5) 4.00 W(10,8) 4.80 W(13,3) 4.20
W(3,11) 4.60 W(7,10) 6.00 W(10,13) 4.80 W(13,8) 3.60
W(3,13) 5.00 W(7,11) 4.40 W(11,3) 4.40 W(13,10) 6.60
W(4,1) 8.80 W(7,13) 3.00 W(11,5) 4.20 W(14,7) 14.2
W(4,3) 4.80 W(8,3) 4.80 W(11,7) 4.60 W(15,2) 8.00
W(5,7) 7.60 W(8,10) 5.40 W(11,12) 6.80
W(5,11) 10.6 W(8,13) 5.20 W(12,3) 4.80
The convergence process of the Lagrangian dual problem is illustrated in

Fig. 3.5. In this figure, Dual UB denotes the current objective value of the
Lagrangian dual function, which can be thought of as an upper bound of the
primal objective value. Primal Feasible Solution denotes the current primal
feasible solution recovered from the Lagrangian dual solution, which can be
thought of as providing a lower bound to the optimal primal objective value.
During each iteration, the cutting-plane method solves the Lagrangian dual
problem. The optimal objective value of the Lagrangian dual problem (or the
upper bound to the primal objective value) is nonincreasing with iterations.
Meanwhile, the best primal feasible objective value keeps increasing with iter-
ations. As expected, the upper bound and the lower bound converge and give
the optimal solution, as shown in this figure. For this 15-node ad hoc network,
the cutting-plane algorithm converged in approximately 115 iterations. The
optimal value of the network utility function is 56.85 (in ln(b/s)).
Table 3.4 Input covariance matrices in the 15-node network (mW).
35.46 0.00 0.00i 12.80 0.03 0.04i

Q(1,4) Q(8,13)
0.00 + 0.00i 35.46 0.03 + 0.04i 12.82
11.82 0.00 0.00i 17.23 0.09 + 0.10i
Q(2,9) Q(9,2)
0.00 + 0.00i 11.82 0.09 0.10i 17.47
24.51 0.08 0.08i 14.78 0.10 0.08i
Q(2,12) Q(9,6)
0.08 + 0.08i 23.98 0.10 + 0.08i 14.55
11.82 0.00 0.00i 17.21 0.32 + 0.28i
Q(2,15) Q(9,15)
0.00 + 0.00i 11.82 0.32 0.28i 16.77
8.90 0.01 + 0.00i 11.82 0.00 0.00i
Q(3,4) Q(10,7)
0.01 0.00i 8.91 0.00 + 0.00i 11.82
7.36 0.10 + 0.01i 11.82 0.0 0.00i
Q(3,8) Q(10,8)
0.10 0.01i 7.23 0.016 + 0.00i 11.82
11.37 0.00 + 0.33i 11.82 0.00 0.00i
Q(3,11) Q(10,13)
0.00 0.33i 12.04 0.00 + 0.00i 11.82
9.60 0.13 0.18i 10.76 0.00 + 0.00i
Q(3,12) Q(11,3)
0.13 + 0.18i 10.27 0.00 0.00i 10.70
12.22 0.35 + 0.69i 10.42 0.01 0.12i
Q(3,13) Q(11,5)
0.35 0.69i 12.09 0.01 + 0.12i 10.39
21.84 0.09 + 0.02i 11.83 0.25 0.01i
Q(4,1) Q(11,7)
0.09 0.02i 21.87 0.25 + 0.01i 12.10
11.82 0.000 0.000i 16.76 0.04 + 0.02i
Q(4,3) Q(11,12)
0.00 + 0.00i 11.82 0.04 0.02i 17.04
11.82 0.00 0.00i 8.87 0.00 0.00i
Q(4,12) Q(12,2)
0.00 + 0.00i 11.82 0.00 + 0.00i 8.87
19.62 0.01 0.08i 11.75 0.00 0.00i
Q(5,7) Q(12,3)
0.01 + 0.08i 19.68 0.00 + 0.00i 11.73
26.56 0.10 0.14i 9.22 0.04 0.09i
Q(5,11) Q(12,4)
0.10 + 0.14i 26.90 0.04 + 0.09i 9.25
21.22 0.00 0.00i 20.85 0.20 + 0.11i
Q(6,9) Q(12,11)
0.00 + 0.00i 21.22 0.20 0.11i 19.47
22.84 0.21 + 0.14i 11.09 0.03 0.05i
Q(6,15) Q(13,3)
0.21 0.14i 23.24 0.03 + 0.05i 11.08
9.54 0.00 0.04i 8.87 0.00 0.00i
Q(7,5) Q(13,7)
0.00 + 0.04i 9.51 0.00 + 0.00i 8.87
14.19 0.05 0.09i 10.00 0.54 0.59i
Q(7,10) Q(13,8)
0.05 + 0.09i 13.82 0.54 + 0.59i 9.69
12.48 1.15 0.43i 15.70 0.15 + 0.01i
Q(7,11) Q(13,10)
1.15 + 0.43i 11.21 0.15 0.01i 15.71
7.54 0.12 0.02i 39.76 0.19 + 0.04i
Q(7,13) Q(14,7)
0.12 + 0.02i 7.51 0.19 0.04i 39.71
7.10 0.00 0.00i 20.58 0.01 + 0.16i
Q(7,14) Q(15,2)
0.00 + 0.00i 7.10 0.01 0.16i 20.50
12.18 0.00 0.00i 11.82 0.00 0.00i
Q(8,3) Q(15,6)
0.00 + 0.00i 12.15 0.00 + 0.00i 11.82
15.00 0.05 0.05i 13.09 0.10 0.08i
Q(8,10) Q(15,9)
0.05 + 0.05i 14.95 0.10 + 0.08i 12.95
Figure 3.5 Convergence 70

behavior of primal and dual
solutions for the 15-node
65 Dual UB
network.
Network utility (ln(b/s))

60
55
50 Primal feasible solution
45
40
0 20 40 60 80 100 120 140 160
Number of iterations
3.8 Chapter summary
Convex programming is a popular and powerful tool to study nonlinear opti-

mization problems. Once a problem is shown to be a convex program, then
there are standard solution techniques and we may even apply directly a solver
to obtain an optimal solution. For many cross-layer convex optimization prob-
lems, the research community is more interested in exploring a solution in the
dual domain. There are two reasons for this approach. First, many cross-layer
problems, once properly formulated in the dual domain, can be decomposed
into subproblems, each of which may be de-coupled from the other layers.
Such a layering-based decomposition offers better insights and interpretations
for the underlying problems. Second, once a problem is decomposed in its
dual domain, the solution may be implemented in a distributed fashion, which
is a highly desirable feature for networking researchers. Although this second
point was not well discussed in this chapter, it was discussed extensively in the
literature and readers are encouraged to explore this on their own (perhaps by
first studying related work [79; 90; 98]).
In this chapter, we illustrated the above points by studying a cross-layer opti-
mization problem for a multi-hop MIMO network. We show that the problem
can be formulated as a convex program. The problem is a cross-layer optimiza-
tion problem as it involves variables at the network, link, and physical layers.
By studying the problem in its dual domain, we showed that the dual problem
can be decomposed into two subproblems: one subproblem solely involving
variables at the network layer and the other problem involving variables at the
link and physical layers. We described how the dual problem can be solved
by a cutting-plane method and how the solution to the primal problem can be
recovered from the solution to the dual problem. By thoroughly understanding

the materials in this chapter, the readers are expected to be equipped with a
powerful optimization tool in networking research.
Appendix: Proof of Theorem 3.1
We first verify that (R, Y, Q, W) is a feasible solution to the primal prob-

lem. To do this, we only need to verify whether or not (3.10) is satisfied
by (R, Y, Q, W). Since the function 1, YT el c(Wl , Ql ) is convex, we
(j )T e c(W (j ) , Q(j ) ))
have that 1, YT el c(Wl , Ql ) K j =0 j (1, Y l l l
0, i.e., (3.10) is satisfied by (R, Y, Q, W). Thus, (R, Y, Q, W) is a feasible
solution to the primal problem.
To show that (R, Y, Q, W) is an optimal solution
to the primal problem, let
F F K (j )
z = f =1 ln rf = f =1 ln j =0 j rf be the primal objective value
corresponding to (R, Y, Q, W) and let z be the optimal objective value of
both the primal and dual problems. Then, we have

F
K F K
K
F
ln j rf j ln(rf ) = j
(j ) (j ) (j )
z = ln(rf )
f =1 j =0 f =1 j =0 j =0 f =1

=z (K)
(u (K)
)z .

The second inequality holds by the concavity of ln(). The fourth equality

holds because j , for 0 j K, is an optimal solution to (3.17) and z(K) is
the optimal objective value of (3.17). The fifth inequality holds by the termi-
nating condition of Algorithm 3.1. The last inequality holds because (u(K) )
is the objective value of the dual problem under a given u(K) , while z is the
optimal objective value of the dual problem among all u values.
On the other hand, z is the objective value of the primal problem achieved
by a feasible solution (R, Y, Q, W), while z is the optimal objective value of
the primal problem. Thus, we have z z .
Combining both results, we have z = z , i.e., (R, Y, Q, W) is an optimal
solution to the primal problem.
3.9 Problems
3.1 Which of the following sets are convex and which are not? Why?
(a) The set in Fig. 3.6.
(b) The set in Fig. 3.7.
(c) {(x1 , x2 ) : x12 + x22 4}.
(d) {(x1 , x2 ) : x1 2x22 0}.
(e) {(x1 , x2 ) : 1 x12 + x22 4}.
59 3.9 Problems
Figure 3.6 Set 1.
Figure 3.7 Set 2.
Figure 3.8 A simple network. (1)

N1 y1 =f1 N3
link 1
link 2 link 3
(1) (2) (1)

y2 = f2 y2 =f3 y3 = f2
N2
3.2 Which of the following functions are convex, concave, affine, or neither?
Why?
(a) f (x) = x 2 + x 4 , x R.
(b) f (x) = x + x 3 , x R.
(c) f (x) = B log2 (1 + x), B is a positive constant and x R+ .
(d) f (x1 , x2 ) = x1 x2 , for x1 , x2 R+ .
(e) f (x1 , x2 ) = 5x1 10x2 , for x1 , x2 R.
3.3 Compared to a nonconvex program, why is it generally easier to obtain
an optimal solution for a convex programming problem?
3.4 Consider a small network shown in Fig. 3.8. Suppose there are two ses-
sions in this network. The first session is from N1 to N3 and includes two
routing paths: one from N1 to N3 directly and the other from N1 to N3 via
N2. The second session is from N1 to N2 and includes the direct routing path
only. Suppose y1(1) = f1 , y2(1) = y3(1) = f2 , y2(2) = f3 . Give the expressions for
A (the node-link incidence matrix), Y (the collection of all flow vectors), and
R (the collection of all source-sink vectors) in the problem formulation.
3.5 In the networking research community, why is the sum of the log func-
tion of session rates a preferred objective function than the sum of rates?
3.6 In the formulation of Problem P in this chapter, which constraints are
associated with the transport, network, link, and physical layers, respectively?
Which constraints are associated with multiple layers?
3.7 The Problem P discussed in this chapter is a convex program. Instead of
solving the problem directly using standard convex optimization approaches,
in this chapter we solve the Lagrangian dual problem and then recover an opti-
mal solution to the original problem. Why is this approach appealing to the
networking research community?
3.8 In the dual problem decomposition, different Lagrangian dual prob-
lems can be devised for the given Problem P, depending on which con-
straints are treated as g(x) 0 and how set X is defined. Why is the constraint
< 1, YT el > c(Wl , Ql ) 0 chosen as g(x) 0?
3.9 In this chapter, a cutting-plane method is employed to solve the
Lagrangian dual problem. Explain the basic idea of this method.
3.10 In Algorithm 3.1, the third step says if zk (uk ), then (z(k) , u(k) ) is
an optimal dual solution. Why?
3.11 Solve the following problem via a dual-based approach:
Minimize x12 ln(x2 )
subject to x1 + 2x2 5 0
x1 , x2 0.
CHAPTER
4 Design of polynomial-time exact

algorithm
When one door of happiness closes, another opens; but often we look so long at
the closed door that we do not see the one which has opened for us.
Helen Keller
4.1 Problem complexity vs. solution complexity
The previous two chapters focus on developing optimal polynomial-time solu-

tions following formal optimization methods from operations research (OR).
For some problems, such an approach may not be always effective, and
could lead to nonpolynomial-time solutions. For these problems, a customized
approach following algorithm design from computer science (CS) could be
more effective and lead to a polynomial-time solution.
It is important to distinguish a (solution) algorithms complexity from the
underlying problems complexity. A problems complexity determines the
potential complexity of any algorithm that is designed to solve this prob-
lem. That is, for a problem not in P, unless P = NP, any algorithm that
can find an optimal solution to this problem must have nonpolynomial-time
complexity. In contrast, if an algorithm (design to solve the problem) has a
nonpolynomial-time complexity, we cannot claim that this problem is not in P.
Another algorithm designed by someone else may well solve the problem with
a polynomial-time complexity.
In this chapter, we illustrate the above approaches and ideas with a case
study. This case study is concerned with an optimal relay node assignment
problem that arises in cooperative communications (CC) [139]. We first for-
mulate the problem as a mixed-integer linear programming (MILP), following
ORs optimization approach. An MILP can be solved optimally, but with a
61
62 Design of polynomial-time exact algorithm
nonpolynomial-time complexity in general. However, this problem is in fact

in P, as we can design an exact algorithm with polynomial-time complexity
following a CS approach.
4.2 Case study: Optimal cooperative relay node assignment
Wireless channels are considered unreliable due to signal fading. Spatial diver-
sity, in the form of employing multiple antennas (i.e., MIMO), has proved
to be very effective in increasing network capacity and reliability. However,
equipping a wireless node with multiple antennas may not always be practical,
as the footprint of multiple antennas may not fit on a wireless node (particularly
a handheld device). In order to achieve spatial diversity without requiring mul-
tiple antennas on the same node, the so-called cooperative communications
(CC) has been introduced [87; 136; 137]. Under CC, each node is equipped
with only a single antenna, and spatial diversity is achieved by exploiting
the antennas on other nodes in the network through cooperative relaying. In
Section 4.3, we give a brief overview of CC so as to set the stage of our case
study.
As expected, the choice of a relay node plays a critical role in the perfor-
mance of CC [18; 22; 180]. An improperly chosen relay node may offer a
smaller data rate for a sourcedestination pair than that under direct trans-
mission. In Section 4.4, we describe the relay node assignment problem in a
network environment, where multiple source-destination pairs compete for the
same pool of relay nodes in the network. Our objective is to assign the avail-
able relay nodes to different sourcedestination pairs so as to maximize the
minimum data rate among all pairs.
In Section 4.5, we will show that this problem can be formulated as a mixed-
integer linear programming (MILP). In general, an MILP has a nonpolyno-
mial complexity. In Section 4.6, instead of studying this MILP, we develop an
optimal polynomial-time algorithm, called ORA, that directly solves the orig-
inal problem. A novel idea in this algorithm is a linear marking mechanism,
which yields a low complexity for each iteration. In Section 4.7, we give a
proof of the optimality of ORA and Section 4.8 presents numerical results.
Section 4.9 summarizes this chapter. The case study in this chapter provides
an interesting instance to show that a complex problem formulation (seemingly
corresponding to a nonpolynomial-time solution) cannot be used as a way to
show that the problem is not in P.
4.3 Cooperative communications: a primer
The essence of CC is best explained by a three-node example in Fig. 4.1. In

this figure, node s is the source node, node d is the destination node, and node
63 4.3 Cooperative communications: a primer
r r r
s d s d s d
(a) One source, one destina- (b) Source node transmits in (c) Relay node transmits in the
tion, and one relay node. the first time slot. second time slot.
Figure 4.1 A three-node relay r is a relay node. Transmission from s to d is done on a frame-by-frame basis.
channel for CC. Within a frame, there are two time slots. In the first time slot, source node s
makes a transmission to the destination node d. Due to the broadcast nature
of wireless communications, this transmission is also overheard by the relay
node r. In the second time slot, node r forwards the data received in the first
time slot to node d. Note that such a two-slot structure is necessary for CC due
to the half-duplex nature of most wireless transceivers.
In this section, we provide expressions for the achievable data rate under CC
and direct transmissions (i.e., no cooperation). For CC, we consider both AF
and DF coding schemes at the relay node [87].
Amplify-and-forward (AF) Let hsd , hsr , and hrd capture the effect of
path-loss, shadowing, and fading between nodes s and d, s and r, and r and
d, respectively. Denote zd [t] and zr [t] as the zero-mean background noise at
nodes d and r in the tth time slot, respectively, with variance d2 and r2 .
Denote xs as the signal transmitted by source node s in the first time slot.
Then the received signal at destination node d, ysd , can be expressed as
ysd = hsd xs + zd [1], (4.1)
and the received signal at the relay node r, ysr , is
ysr = hsr xs + zr [1]. (4.2)
In the second time slot, relay node r transmits to destination node d. The
received signal at d, yrd , can be expressed as
yrd = hrd r ysr + zd [2],
where r is the amplifying factor at relay node r and ysr is given in (4.2). Thus,
we have
yrd = hrd r (hsr xs + zr [1]) + zd [2]. (4.3)
The amplifying factor r at relay node r should satisfy the power constraint
r2 (|hsr |2 Ps + r2 ) = Pr , where Ps and Pr are the transmission powers at
nodes s and r, respectively. So, r is given by
Pr
r2 = .
|hsr |2 Ps + r2
We can rewrite (4.1), (4.2), and (4.3) in the following compact matrix form:
Y = Hxs + BZ,
where

zr [1]
ysd hsd 0 10
Y= , H= , B= , and Z = zd [1] .
yrd r hrd hsr r hrd 01
zd [2]
(4.4)
It has been shown in [87] that the above channel, which combines both the
direct path (s to d) and the relay path (s to r to d), can be modeled as a
one-input, two-output complex Gaussian noise channel. The achievable data
rate CAF (s, r, d) from s to d is given by
W ! "
CAF (s, r, d) = log2 det I + (Ps HH )(BE[ZZ ]B )1 , (4.5)
2
where W is the bandwidth, det() is the determinant function, I is the identity
matrix, the superscript represents the complex conjugate transpose, and
E[] is the expectation function.
Substituting (4.4) into (4.5) and performing algebraic manipulations, we
have
# $
W Ps Ps |hsr |2 Pr |hrd |2
CAF (s, r, d) = log2 1 + 2 |hsd | +
2
.
2 d Ps d2 |hsr |2 + Pr r2 |hrd |2 + r2 d2
(4.6)
Denote
Ps Ps Pr
SNRsd = |hsd |2 , SNRsr = 2 |hsr |2 , and SNRrd = 2 |hrd |2 .
d2 r d
We have
CAF (s, r, d) = W IAF (SNRsd , SNRsr , SNRrd ),
where
% &
1 SNRsr SNRrd
IAF (SNRsd , SNRsr , SNRrd ) = log2 1 + SNRsd + .
2 SNRsr + SNRrd + 1
Decode-and-forward (DF) Under this mode, relay node r decodes and

estimates the received signal from source node s in the first time slot, then
transmits the estimated data to destination node d in the second time slot. The
achievable data rate for DF under the two time-slot structure is given by [87]:
CDF (s, r, d) = W IDF (SNRsd , SNRsr , SNRrd ),
where
65 4.4 The relay node assignment problem
IDF (SNRsd , SNRsr , SNRrd )

1
= min{log2 (1 + SNRsr ), log2 (1 + SNRsd + SNRrd )}.
2
Note that IAF () and IDF () are increasing functions of Ps and Pr , respec-
tively. This suggests that, in order to achieve the maximum data rate under
either mode, both source node and relay node should transmit at their maxi-
mum powers. In this chapter, we let Ps = Pr = P .
Direct transmission When CC is not used, source node s transmits to desti-
nation node d in both time slots. The achievable data rate from node s to node
d is
CD (s, d) = W log2 (1 + SNRsd ).
Based on the above results, we have two observations. First, comparing CAF
(or CDF ) to CD , we can see that CC may not always be better than direct trans-
missions. For example, a poor choice of relay node could make the achievable
data rate under CC lower than that under direct transmissions. This fact under-
lines the significance of relay node selection in CC. Second, although AF and
DF are different mechanisms, the data rates for both of them have the same
form, i.e., a function of SNRsd , SNRsr , and SNRrd . Therefore, we only need
to develop one optimal relay node assignment algorithm, which should work
for both AF and DF. Table 4.1 lists the notations used in this chapter.
4.4 The relay node assignment problem
Based on the background in the last section, we consider a relay node assign-
ment problem in a network setting. There are N nodes in an ad hoc network,
with each node being either a source node, a destination node, or a poten-
tial relay node (see Fig. 4.2). In order to avoid interference, we assume that
orthogonal channels are available in the network (e.g., using OFDMA) for
CC [87]. The path-loss between nodes u and v is captured in huv and is
given a priori. Denote Ns = {s1 , s2 , . . . , sNs } as the set of source nodes, Nd =
{d1 , d2 , . . . , dNd } as the set of destination nodes, and Nr = {r1 , r2 , . . . , rNr } as
the set of relays (see Fig. 4.2). We consider unicast where every source node
si is paired with a destination node di , i.e., Nd = Ns . Each node is equipped
with a single transceiver and can transmit/receive within one channel at a time.
Further, we assume that each node can only serve a unique role of source,
destination, or relay. That is, Nr = N 2Ns .
Note that a source node may not always get a relay node. There are two
possible scenarios in which this may happen. First, there may not be a sufficient
number of relay nodes in the network (e.g., Nr < Ns ). In this case, some source
nodes will not have relay nodes. Second, even if there are enough relay nodes,
Table 4.1 Notation.
Symbol Definition
Asi ,rj Decision variable for relay node assignment

CR (si , rj ) Achievable rate for si di pair when relay node rj is used
CR (si , ) Achievable rate for si di pair under direct transmission
Cmin The minimum rate among all sourcedestination pairs
huv Effect of path-loss, shadowing, and fading from node u to node v
Ns Set of source nodes in the network
Nr Set of relay nodes in the network
Ns = |Ns |, number of source nodes in the network
Nr = |Nr |, number of relay nodes in the network
N Number of all the nodes in the network
P Maximum transmission power
rj The j th relay node, rj Nr
R (si ) The relay node assigned to si under
si The ith source node, si Ns
S (rj ) The source node that uses rj under
SNRuv The signal-to-noise ratio between nodes u and v
W Channel bandwidth
xs Signal transmitted by node s
yuv Received signal at node v (form node u)
zv [t] Background noise at node v in the tth time slot
r Amplifying factor at relay r
v2 Variance of background noise at node v
A solution for relay node assignment
Figure 4.2 An ad hoc network Sender Receiver Potential relay node

consisting of source nodes
(senders), destination nodes
(receivers), and relay nodes.
a sender may choose not to use a relay node if it leads to a lower data rate than
direct transmission.
We now discuss the objective function of our problem. Although different
objectives can be used, a widely used objective for CC is the achievable data
rate. For the multi-session network environment considered in this chapter
(see Fig. 4.2), each sourcedestination pair will have a different achievable
data rate after we apply a relay node assignment algorithm. So a plausible
objective is to maximize the minimum achievable data rate among all the
sourcedestination pairs.
More formally, denote R(si ) as the relay node assigned to si , and S(rj ) as
the source node that uses rj . For both AF and DF, its achievable data rate can
be written as (see Section 4.3)
67 4.5 An optimization-based formulation
W IR (SNRsi ,di , SNRsi ,R(si ) , SNRR(si ),di ),
with IR () = IAF () when AF is employed, and IR () = IDF () when DF is

employed. In case si does not use a relay, we denote R(si ) = and the
achievable data rate is the direct transmission capacity, i.e.,
CR (si , ) = CD (si , di ).
Combining both cases, we have

'
W IR (SNRsi ,di , SNRsi ,R(si ) , SNRR(si ),di ) if R(si ) = ,
CR (si , R(si )) =
W log(1 + SNRsi ,di ) if R(si ) = .
(4.7)
Note that di is not listed in function CR (si , R(si )) since for each source node
si , its corresponding destination node di is deterministic.
Denote by Cmin the objective function, which is the minimum achievable
data rate among all source nodes. We have
Cmin = min{CR (si , R(si )) : si Ns }.
Our goal is to find an optimal relay node assignment for all the source
destination pairs such that Cmin is maximized.
4.5 An optimization-based formulation
In this section, we present a formulation for the relay node assignment problem
based on an optimization approach from OR. To describe relay node assign-
ment mathematically, we denote a binary variable Asi ,rj as
'
1 if relay rj is assigned to source si ,
Asi ,rj =
0 otherwise.
Since a source node si will be assigned at most one relay node, we have

Nr
Asi ,rj 1.
j =1
Due to the unique role of each node, a relay node rj can be assigned to at most
one source node. That is

Ns
Asi ,rj 1.
i=1
Table 4.2 A mathematical formulation for relay node assignment problem.
Maximize Cmin

Nr
subject to Cmin 1 Asi ,rj CR (si , )
j =1

Nr
+ Asi ,rj CR (si , rj ) (si Ns )
j =1

Ns
Asi ,rj 1 (rj Nr )
i=1
Nr
Asi ,rj 1 (si Ns )
j =1
Asi ,rj {0, 1}, Cmin 0 (si Ns , rj Nr )
Note that if any Asi ,rj = 1, then a relay rj is assigned to source si . By (4.7),
the achievable data rate from source si to destination di by using relay rj
r
is CR (si , rj ). On the other hand, if N j =1 Asi ,rj = 0, then there is no relay
assigned to source si . By (4.7), the achievable data rate from source si to des-
tination di is CR (si , ). Combining these two cases, the achievable data rate
from source si to destination di is

Nr
Nr
1 Asi ,rj CR (si , ) + Asi ,rj CR (si , rj ).
j =1 j =1
Based on the above discussion, one formulation for the optimal relay node
assignment problem is given in Table 4.2. This formulation is a mixed-integer
linear programming (MILP) and is NP-hard in general [46]. As discussed in
Section 4.1, this does not mean that our optimal relay node assignment prob-
lem is NP-hard because an NP-hard formulation is not a proof of a problems
NP-hardness. In fact, this problem can be viewed as a maximin linear assign-
ment problem, for which there exists polynomial-time algorithms (see [3],
[25], [89], [126], and [161] for example, although these papers address dif-
ferently structured problems, the proposed algorithms can be adapted to solve
our problem in polynomial time as well).
Instead of applying a general algorithm for the maximin linear assignment
problem, we design in the next section a specialized polynomial-time algo-
rithm (along with an optimality proof) for our relay node assignment problem.
The existence of such a polynomial-time algorithm implies that the problem
belongs to the class P of polynomially solvable problems.
69 4.6 An exact algorithm
4.6 An exact algorithm

4.6.1 Basic idea
The polynomial-time algorithm that we will describe is called the Optimal
Relay Assignment (ORA) algorithm. Fig. 4.3 shows the flow chart of the ORA
algorithm.
Initially, the ORA algorithm starts with a random but feasible relay node
assignment. By feasible, we mean that each sourcedestination pair can be
assigned at most one relay node and that a relay node can be assigned only
once. Such an initial feasible assignment is easy to construct, e.g., a direct
transmission between each sourcedestination pair (without the use of a relay)
is a special case of a feasible assignment.
Starting with this initial assignment, ORA adjusts the assignment during
each iteration, with the goal of increasing the objective function Cmin . Specif-
ically, during each iteration, ORA identifies the source node that corresponds
to the current Cmin . Then, ORA helps this source node search for a better
relay such that this bottleneck data rate can be increased. In the case that
the selected relay is already assigned to another source node, a further adjust-
ment of the relay node assignment for that source node is necessary (so that its
current relay can be released). Such an adjustment might have a chain effect
on a number of source nodes in the network. It is important that for any adjust-
ment made on a relay node assignment, the affected source node still maintains
a data rate larger than Cmin . There are only two outcomes from such an iter-
ation: (i) a better assignment is found, in which case ORA moves on to the
next iteration; or (ii) a better assignment cannot be found, in which case ORA
terminates.
There are two key issues that we must address in the design of this algorithm.
First, for any nonoptimal solution, the algorithm should be able to find a better
solution. As a result, upon termination the final assignment is optimal. Sec-
ond, its running time must be polynomial. We will show that ORA addresses
both issues successfully. Specifically, we show that the complexity of the ORA
algorithm is polynomial in Section 4.6.4. We will also give a correctness proof
of its optimality in Section 4.7.
4.6.2 Algorithmic details

To begin with, the ORA algorithm performs a preprocessing step. In this
step, for each sourcedestination pair, the source node si considers each relay
node rj in the network and computes the corresponding data rate CR (si , rj ) by
(4.7). Each source node si also computes the achievable data rate CR (si , ) by
(4.7) under direct transmissions (i.e., without the use of a relay node). After
these computations, each source node si can identify those relay nodes that
Figure 4.3 A flow chart of the

BEGIN
ORA algorithm.
Preprocessing, and Start the search

Initial relay assignment
Identify the source sb

with minimum data rate Cmin.
YES
Can we find
a better solution?
Use Find_Another_Relay(sb) to
improve the data rate of sb,
NO and return the outcome.
END
Find_Another_Relay(si )
BEGIN
Can we find an
Return NO unmarked relay for si
NO with data rate larger
than Cmin?
YES
NO Is this relay
already assigned?
YES
Mark this relay, and denote

its corresponding source as sj
For sj , use
YES Find_Another_Relay(sj) to NO
Better solution found. determine if another
relay can be
assigned
Clear marks on
all relays.
Return
YES
Figure 4.4 An example tree r1 s1

topology in the ORA algorithm
for finding a better solution.
r2 s2 r4 s4
r3 s3 r5 s5 r6 s6
cannot find cannot find

another relay another relay r7
can be
assigned to s6
can offer an increase in its data rate compared to direct transmissions, i.e.,
those relays with CR (si , rj ) > CR (si , ). Obviously, it only makes sense to
consider these relays for CC. That is, for a source node si , those relay nodes
with rates not greater than direct transmissions can be removed from further
consideration.
After the preprocessing step, we enter the initial assignment step. The objec-
tive of this step is to obtain an initial feasible solution for the ORA algorithm
so that it can start its iteration. In the preprocessing step, we have already iden-
tified the list of relay nodes for each source node that can increase its data rate
compared to direct transmission. We can randomly assign a relay node from
this list to a source node. Note that once a relay node is assigned to a source
node, it cannot be assigned again to another source node. If there is no relay
node available for a source node, then this source node will simply employ
direct transmission. Upon the completion of this assignment, each source node
will have a data rate that is no less than that under direct transmission.
The next step in the ORA algorithm is to find a better assignment, which
is an iterative process. This is the key step in the ORA algorithm. The detail
of this step is shown in the bottom portion of Fig. 4.3. As a starting point of
this step, the ORA algorithm identifies the smallest data rate Cmin among all
sources. The ORA algorithm aims to increase this minimum data rate for the
corresponding source node, while having all other source nodes maintain their
data rates above Cmin . Without loss of generality, we use Fig. 4.4 to illustrate
the search process.
Suppose ORA identifies that s1 has the smallest rate Cmin under the current
assignment (with relay node r1 ). In case of a tie, i.e., when two or more
source nodes have the same smallest data rate, the tie is broken by choosing
the source node with the highest node index. Source s1 examines other relays
with a rate larger than Cmin . If it cannot find such a relay, then no better
solution is found and the ORA algorithm terminates.
Otherwise, i.e., when there exist better relays, we consider these relays in a
non-increasing order of data rate (should it be assigned to s1 ). That is, we try
the relay that can offer the maximum possible increase in data rate first. In
case of a tie, i.e., when two or more relay nodes offer the same maximum data
rate, the tie is broken by choosing the relay node with the highest node index.
Suppose that source node s1 considers relay node r2 . If this relay node is not
yet assigned to any other source node, then r2 can be immediately assigned
to s1 . In this simple case, we find a better solution and the current iteration
is completed.
Otherwise, i.e., if r2 is already assigned to a source node, say s2 , we mark
r2 to indicate that r2 is under consideration and check whether r2 can be
released by s2 .
To release r2 , source node s2 needs to find another relay (or use direct trans-
mission) while making sure that such a new assignment still keeps its data
rate larger than the current Cmin . This process is identical to that for s1 , with
the only (but important) difference being that s2 will not consider a relay
that has already been marked, because that relay node has already been
considered by a source node encountered earlier in this iteration.
Suppose that source node s2 now considers relay r3 . We consider the follow-
ing three cases. If this relay node is not yet assigned to any source node, then
r3 can be assigned to s2 , r2 can be assigned to s1 , and the current iteration is
completed. Moreover, if the relay under consideration by s2 is the one that is
being used by the source node that initiated the iteration, i.e., relay r1 , then
it is easy to see that r1 can be taken away from s1 . A better solution (r1 is
assigned to s2 and r2 is assigned to s1 ) is found and the current iteration is
completed. Otherwise, we mark r3 and check further to see whether r3 can
be released by its corresponding source node, say s3 . We also note that s2
can consider direct transmission if it offers a data rate larger than Cmin .
Suppose that s3 cannot find any unmarked relay that offers a data rate
larger than Cmin and its achievable data rate under direct transmission is not
larger than Cmin . Then s2 cannot use r3 as its relay.
If any unmarked relay that offers a data rate larger than Cmin cannot be
assigned to s2 , then s1 cannot use r2 and the algorithm will move on to con-
sider the next relay on its nonincreasing rate list, say r4 .
The search continues, with relay nodes being marked along the way, until a
better solution is found or no better solution can be found. For example,
in Fig. 4.4, s6 finds a new relay r7 . As a result, we have a new assign-
ment, where r7 is assigned to s6 , r6 is assigned to s4 , and r4 is assigned
to s1 .
Note that the mark on a relay node will not be cleared throughout the
search process in the same iteration. We call this a linear marking mech-
anism. These marks will only be cleared when the current iteration termi-
nates and before the start of the next iteration. A pseudocode for the ORA
Table 4.3 An example illustrating the operation of the ORA algorithm.
(a) After initial assignment. The start of the first iteration.
r1 r2 r3 r4 r5 r6 r7
s1 14 7 24 5 14 15 17 9
s2 9 8 10 11 20 10 12 11
s3 11 10 13 17 21 8 9 19
s4 12 8 9 12 11 10 9 18
s5 10 9 18 19 24 9 13 23
s6 7 18 12 6 11 11 17 20
s7 16 1 9 4 14 19 8 12
(b) The start of the second iteration.
s1 14 7 24 5 14 15 17 9
s2 9 8 10 11 20 10 12 11
s3 11 10 13 17 21 8 9 19
s4 12 8 9 12 11 10 9 18
s5 10 9 18 19 24 9 13 23
s6 7 18 12 6 11 11 17 20
s7 16 1 9 4 14 19 8 12
(c) The start of the third iteration.
s1 14 7 24 5 14 15 17 9
s2 9 8 10 11 20 10 12 11
s3 11 10 13 17 21 8 9 19
s4 12 8 9 12 11 10 9 18
s5 10 9 18 19 24 9 13 23
s6 7 18 12 6 11 11 17 20
s7 16 1 9 4 14 19 8 12
(d) Final assignment upon ORA termination.
s1 14 7 24 5 14 15 17 9
s2 9 8 10 11 20 10 12 11
s3 11 10 13 17 21 8 9 19
s4 12 8 9 12 11 10 9 18
s5 10 9 18 19 24 9 13 23
s6 7 18 12 6 11 11 17 20
s7 16 1 9 4 14 19 8 12
algorithm is shown in Algorithm 4.1. Note that ORA works regardless of

whether Nr Ns or Nr < Ns .
We now use an example to illustrate the operation of the ORA algorithm, in
particular, its linear marking mechanism.
Example 4.1
Suppose that there are seven sourcedestination pairs and seven relay nodes
in the network.
Algorithm 4.1 The ORA algorithm

Main algorithm
1. Calculate CR (si , ) and CR (si , rj ) for each source node si and relay
node rj .
2. For each source node si , build a list of relay nodes following
nonincreasing order of rates (including CR (si , )).
3. Find an initial relay node assignment.
4. while (true) {
5. Set all relay nodes in the network as unmarked.
6. Denote sb as the bottleneck source node with rate Cmin .
Its relay node is R(sb ).
7. Find Another Relay(sb , R(sb ), Cmin , sb ).
8. If sb cannot find another relay
9. break }
Subroutines
Find Another Relay(si , rj , Cmin , sb ) {
10. Search si s relay node list {
11. Suppose the current considered relay node is rk .
12. If CR (si , rk ) Cmin
13. break.
14. If relay rk is marked
15. continue.
16. Check Relay Availability(rk , Cmin , sb )
17. If rk is available {
18. Remove relay node rj s assignment to si .
19. Assign relay node rk to si .
20. return si finds another relay. } }
21. return si cannot find another relay }
Check Relay Availability(rj , Cmin , sb ) {
22. If rj = , rj = R(sb ), or rj is not assigned to any source node
23. return rj is available.
24. Set rj as marked.
25. Find Another Relay(S(rj ), rj , Cmin , sb )
26. If S(rj ) can find another relay
27. return rj is available.
28. else
29. return rj is unavailable. }
Note: A tie is broken by choosing the node with the highest node index.
Table 4.3(a) shows the data rate for each source node si when relay node
rj is assigned to it. The symbol indicates direct transmission. Also shown
in Table 4.3(a) is an initial relay node assignment, which is indicated by an
underscore on the intersecting row (si ) and column (rj ). That is, the initial
assignment is r5 for s1 , r4 for s2 , r2 for s3 , r7 for s4 , r3 for s5 , r6 for s6 , and
for s7 . Note that the initial assignment step ensures that the achievable
data rate for each sourcedestination pair is no less than that under direct
transmission.
Given the initial relay node assignment in Table 4.3(a), source s3 is iden-
tified as the bottleneck source node sb , with the current Cmin = 13. Since
consideration of relay nodes is given in the order of nonincreasing (from
largest to smallest) data rate, r4 is therefore considered for s3 first. Since r4
is already assigned to source node s2 , we mark r4 now. Next we check
whether or not s2 can find another relay. But any other relay (or direct trans-
mission) will result in a data rate no greater than the current Cmin = 13. This
means that r4 cannot be taken away from s2 . Therefore, s3 will then con-
sider the next relay node that offers the second largest rate value, i.e., r7 .
Since r7 is already assigned to sender s4 , we mark r7 . Next, the ORA
algorithm will check whether or not s4 can find another relay. But none of
the relay nodes except r7 can offer a data rate larger than the current Cmin to
s4 . So r7 cannot be taken away from s4 . Therefore, node s3 will now check
for the relay node that offers the next highest rate, i.e., r3 . Since r3 is already
assigned to s5 , we mark r3 now. Next, the ORA algorithm checks to see
if s5 can find another relay; s5 checks relay nodes in nonincreasing order of
data rate values. Since r4 (with largest rate) and r7 (with the second largest
rate) are both marked, they will not be considered (this is the essence of
the linear marking mechanism of the ORA algorithm, which ensures that
the linear complexity of examining relay nodes in each iteration). The next
relay to be considered is r2 , which offers a rate of 18 > Cmin = 13. More-
over, r2 is the relay node assigned to sb = s3 . Thus, r2 can be reassigned to
s5 and r3 can be reassigned to s3 . The new assignment after the first itera-
tion is shown in Table 4.3(b). Now the objective value, Cmin , is updated to
15, which corresponds to s1 . Before the second iteration, all markings done
in the first iteration are cleared.
In the second iteration, the ORA algorithm will identify s1 as the source
node with the minimum data rate in the network. The algorithm will then
perform a new search for a better relay node for source s1 . Similar to the
first iteration, the assignments for other source nodes can change during
this search process, but all assignments should result in data rates larger
than 15. The details of this iteration are similar to those in the first iteration
and we leave them as a homework exercise.
Similar to the previous two iterations, in the third iteration, the ORA
algorithm will identify the source node with minimum data rate, which is
now s7 . The algorithm will then perform a new search to improve the data
rate of s7 .
Finally, in the fourth and final iteration, the algorithm will try to improve
the data rate of s3 and will be unsuccessful. As a result, the ORA algorithm
will terminate.
The final assignment upon termination of the ORA algorithm is shown
in Table 4.3(d), with the optimal (maximum) value of Cmin being 17.
4.6.3 Caveat on the marking mechanism

We now re-visit the marking mechanism in the ORA algorithm. Although dif-
ferent marking mechanisms may be designed to achieve an optimal solution,
their complexities may differ significantly. In this subsection, we first present a
naive marking mechanism, which appears to be natural approach but unfortu-
nately leads to an exponential complexity for each iteration. Then we discuss
our marking mechanism and show that it leads to a linear complexity of exam-
ining relay nodes in each iteration.
A natural and naive approach is to perform unmarking during an iteration.
This approach is best explained with an example. Again, lets look at Fig. 4.4.
Source node s1 first considers r2 . Since r2 is now being considered by s1 and
is used by s2 in the current solution, r2 is marked. Source node s2 considers r3 ,
which is already assigned to s3 . Since s3 cannot release r3 without reducing its
data rate below the current Cmin , this branch of search is futile and s1 moves on
to consider its next candidate relay node r4 . Since r4 is currently assigned to s4 ,
we mark r4 and try to find a new relay for s4 . Now the question is: should those
marks on r2 and r3 that we put earlier in this iteration be removed so that both
of them can be considered by s4 ? Under this naive approach, r2 and r3 should
be unmarked so that they can be considered as candidate relay nodes for s4
in its search. Similarly, when we try to find a relay for s6 , relay nodes r2 , r3 ,
r4 , and r5 should be unmarked so that they can be considered as candidate
relay nodes for s6 , in addition to r7 . In summary, under this approach, each
relay node that has been considered earlier in the search process by a source
node should be unmarked when this source node moves on to consider the
next relay node, so that this relay node can be in the pool of candidate relay
nodes to be considered in the search process. It is not hard to show that such
a marking/unmarking mechanism will consider all possible assignments and
can guarantee to find an optimal solution upon termination. However, it can be
shown that the complexity of this approach is exponential for each iteration.
We leave it as a homework exercise.
In contrast, under the ORA algorithm, unmarking of marked nodes is only
performed at the end of an iteration (so that there is a clean start for the next
iteration). That is, marked relay nodes in the search process will remain marked
throughout the iteration. For example, in Fig. 4.4, when s4 tries to find another
relay, it will no longer consider r2 and r3 that have already been marked.
Similarly, when s6 tries to find another relay, it will not consider r2 , r3 , r4 , and
r5 . As a result, any relay node will be considered at most once in the search
process and the pool of candidate relay nodes shrinks linearly as the search
continues. It is not hard to see that this marking mechanism leads to a linear
complexity of examining relay nodes in each iteration of the ORA algorithm.
Although such linear marking mechanism is attractive in its complexity, a
natural question we would ask is: How could such a linear marking lead to
an optimal solution, as it may overlook many possible assignments that poten-
tially increase Cmin ? This is precisely the question that we will address in
Section 4.7, where we will prove that ORA can guarantee that its final solution
is optimal (Theorem 4.1).
Remark 4.1
We comment here that the linear marking mechanism can be alternatively
viewed as a flow augmentation step within a specially defined max-flow
problem (see Problem 4.12 in the exercise).
4.6.4 Complexity analysis

We now analyze the computational complexity of the ORA algorithm in Algo-
rithm 4.1.
Before first iteration The complexity of Step 1 is O(Ns (Nr + 1)) =
O(Ns Nr ). The complexity of Step 2 is O(Ns (Nr + 1) log2 (Nr + 1)) =
O(Ns Nr log2 Nr ). The complexity of Step 3 is O(Ns ).
Each iteration The dominant component in complexity is the recursive pro-
cess of searching an unmarked relay node and checking its availability. This
process may involve all source nodes, each of which checking its relay nodes
marking status and availability. Since a source node needs to check at most Nr
relay nodes and the number of source nodes is Ns , the complexity of checking
relay nodes marking status and availability in an iteration is O(Ns Nr ). Given
this is the dominant complexity in an iteration, we have that the complexity of
an iteration is O(Ns Nr ).
Number of iterations We examine the maximum number of iterations that
the ORA can execute. Note that the number of different Cmin values that we
may encounter is no more than the number of sourcerelay pairs and that the
Cmin value for the current bottleneck node sb is increased after each improving
iteration. As a result, the number of iterations that the algorithm can go through
is O(Ns Nr ).
Complexity of all iterations The complexity of all iterations is O(Ns Nr )
O(Ns Nr ) = O(Ns2 Nr2 ).
Overall complexity Thus, the overall complexity of ORA algorithm

can be summed up as O(Ns Nr ) + O(Ns Nr log2 Nr ) + O(Ns )+ O(Ns2 Nr2 ) =
O(Ns2 Nr2 ).
4.7 Proof of optimality
In this section, we give a correctness proof of the ORA algorithm. That is,
upon termination of the ORA algorithm, the solution (i.e., objective value and
the corresponding relay node assignment) is optimal.
Our proof is based on contradiction. Denote as the final solution obtained
by the ORA algorithm, with the objective value being Cmin . For , denote the
relay node assigned to source node si as R (si ). Conversely, for , denote the
source node that uses relay node rj as S (rj ).
Assume that there exists a better solution than . That is, the objective
value for , denoted as Cmin , is greater than that for , i.e., Cmin > Cmin . For
, denote the relay node assigned to source node si as R (si ), the source node
that uses relay node rj as S (rj ).
The key idea in the proof is to exploit the marking status of the ORA algo-
rithm at the end of the last iteration, which is a nonimproving iteration. Specif-
ically, in the beginning of this last iteration, ORA will select a bottleneck
source node, which we denote as sb . ORA will then try to improve the solution
by searching for a better relay node for this bottleneck source node. Since the
last iteration is a nonimproving iteration, ORA will not find a better solution,
and thus will terminate. We will show that R (sb ) is not marked at the end of
the last iteration of ORA. On the other hand, if there exists a better solution
than , we will show that R (sb ) is marked at the end of the last iteration of
ORA. This leads to a contradiction and thus the assumption of the existence of
a better solution cannot hold. We begin our proof with the following fact:
Fact 4.1
For the bottleneck source node sb under , its relay node R (sb ) is not
marked at the end of the last iteration of the ORA algorithm.
Proof. In ORA algorithm, a relay node rj is marked only if rj = ,

rj = R (sb ), and rj is assigned to some source node (see Check Relay
Availability() in Algorithm 4.1). Thus, R (sb ) cannot be marked at the end
of the last iteration of the ORA algorithm.
Fact 4.1 will be the basis for contradiction in our proof of Theorem 4.1, the
main result of this section. Now we present the following three claims, which
recursively examine relay node assignments under . First, for the relay node
assigned to sb in , i.e., R (sb ), we have:
79 4.7 Proof of optimality
Figure 4.5 The sequence of R(sb) sb

nodes under analysis in the ^ R^(sb)
proof of optimality. (marked)
(unmarked)
G(sb) ^ R^( G(sb) )
(marked)

^ R^( Gk(sb) )
Gk(sb)
(marked)

^ R^( Gn(sb) )
Gn(sb)
(marked)
Claim 4.1
Relay node R (sb ) must be marked at the end of the last iteration of the
ORA algorithm. Further, it cannot be and must be assigned to some source
node under solution .
Proof. Since is a better solution than , we have CR (sb , R (sb )) Cmin >
Cmin . Thus, by construction, ORA will consider the availability of relay node
R (sb ) for sb in its last iteration of . Since the ORA algorithm cannot find a
better solution in this last iteration, R (sb ) must be marked and then the out-
come of checking the availability of R (sb ) shows that it is unavailable. By
the linear marking process, the mark on R (sb ) will not be cleared through-
out the search process in the last iteration. Thus, R (sb ) is marked at the end
of the last iteration of the ORA algorithm.
We now prove the second statement by contradiction. If R (sb ) were
, then sb would choose in the last iteration of ORA since it can offer
CR (sb , R (sb )) > Cmin . But this contradicts the fact that we are now in the
last iteration of ORA, which is a nonimproving iteration. So R (sb ) cannot be
. Further, since we proved that R (sb ) is marked at the end of the last iteration
of the ORA algorithm, it must be assigned to some source node already.
By the definition of S (), we have that R (sb ) is assigned to source

node S (R (sb )) in solution . To simplify the notation, define the function
G () as
G () = S (R ()). (4.8)
Thus, relay node R (sb ) is assigned to source node G (sb ) in (see Fig. 4.5).
Since R (sb ) = R (sb ), they are assigned to different source nodes in ,
i.e., G (sb ) = sb . Now we recursively investigate the relay node assigned to
G (sb ) under the solution , i.e., R (G (sb )). We have the following claim
(also see Fig. 4.5):
Claim 4.2
Relay node R (G (sb )) must be marked at the end of the last iteration of
the ORA algorithm. Further, it cannot be and must be assigned to some
source node under the solution .
The proofs for both statements in this claim follow the same token as that
for Claim 4.1 and are thus left as a homework exercise.
Again, by the definition of S (), we have that the relay node R (G (sb ))
is assigned to source node S (R (G (sb ))) in solution . By (4.8), we have
source node S (R (G (sb ))) = G (G (sb )). To simplify the notation, we
define the function G2 () as
G2 () = G (G ()).
Thus, the relay node R (G (sb )) is assigned to source node G2 (sb ) in .

Now we have two cases: the source node G2 (sb ) may or may not be a node
in {sb , G (sb )}. If the source node G2 (sb ) is a node in {sb , G (sb )}, then we
terminate our recursive procedure. Otherwise, we continue to consider its relay
node in .
In general, we use the following notation:
G0 (sb ) = sb
Gk (sb ) = G (Gk1 (sb )) (k = 1, 2, . . .). (4.9)
Since the number of source nodes is finite, our recursive procedure will termi-
nate in a finite number of steps. Suppose that we terminate at k = n.
Following the same token for Claims 4.1 and 4.2, we can obtain
a similar claim for each of the relay nodes R (G2 (sb )), R (G3 (sb )),
. . . , R (Gk (sb )), . . . , R (Gn (sb )) (see Fig. 4.5). Thus, we can generalize the
statements in Claims 4.1 and 4.2 for relay node R (Gk (sb )) and have the fol-
lowing claim:
Claim 4.3
Relay node R (Gk (sb )) must be marked at the end of the last iteration of
the ORA algorithm. Further, it cannot be and must be assigned to some
source node under solution , k = 0, 1, 2, . . . , n.
Proof. Since is a better solution than , we have CR (Gk (sb ),

R (Gk (sb ))) Cmin > Cmin . Note that Gk (sb ) is some source node in the
81 4.7 Proof of optimality
solution obtained by ORA, whereas R (Gk (sb )) is the relay node assigned
to this source node in the hypothetical better solution . Our goal is to show
that ORA should have marked this relay node in its last iteration.
Since CR (Gk (sb ), R (Gk (sb ))) > Cmin and R (Gk (sb )) is not assigned to
G (sb ) in the last iteration of ORA, then by construction of ORA, ORA must
k
have checked the availability of R (Gk (sb )) for Gk (sb ) during the last iter-
ation, then marked it, and then determined that it is unavailable for Gk (sb ).
Moreover, due to the linear marking process, this mark on R (Gk (sb ))
should remain after the last iteration of ORA. Thus, we conclude that
R (Gk (sb )) is marked at the end of the last iteration of the ORA algorithm.
We now prove the second statement by contradiction. If R (Gk (sb ))
is , then Gk (sb ) will choose in the last iteration since it can offer
CR (Gk (sb ), R (Gk (sb )) > Cmin , and finally sb will be able to get a better
relay node. But this contradicts the fact that this last iteration is a nonim-
proving iteration. So, R (Gk (sb )) cannot be . Further, since we proved that
R (Gk (sb )) is marked at the end of the last iteration of the ORA algorithm,
it must be assigned to some source node already.
Referring to Fig. 4.5, we have Claim 4.3 for a set of relay nodes
R (sb ), R (G (sb )), . . . , R (Gn (sb )). Our recursive procedure terminates
at R (Gn (sb )) because its assigned source node in solution is a node in
{sb , G (sb ), . . . , Gn (sb )}. We are now ready to prove the following theorem,
which is the main result of this section:
Theorem 4.1
Upon termination of the ORA algorithm, the obtained solution is optimal.
Proof. Suppose that the solution obtained by the ORA algorithm is not opti-
mal. Then, there exists a better solution . We have Fact 4.1 and Claim 4.3.
We will show a contradiction.
Under Claim 4.3, we proved that the relay node R (Gn (sb )) is assigned to a
source node in the solution obtained by ORA. Since our recursive procedure
terminates at R (Gn (sb )), its assigned source node in solution is a node
in {sb , G (sb ), . . . , Gn (sb )}. But we also know that under , the source nodes
G (sb ), G2 (sb ), G3 (sb ), . . . , Gn (sb ) have relay nodes R (sb ), R (G (sb )),
R (G2 (sb )), . . . , R (Gn1 (sb )), respectively. Thus, R (Gn (sb )) can only be
assigned to sb in solution . On the other hand, the relay node assigned to sb in
the solution is denoted by R (sb ). Therefore, we must have R (Gn (sb )) =
R (sb ).
Now, Claim 4.3 states that R (Gn (sb )) must be marked after the last
iteration, whereas Fact 4.1 states that the relay node assigned to the bottle-
neck source node, i.e., R (sb ), cannot be marked. Since both R (sb ) and
R (Gn (sb )) are the same relay node, we have a contradiction. Thus, our
assumption that there exists a better solution than does not hold.
Note that the proof of Theorem 4.1 does not depend on the initial assign-
ment in the ORA algorithm. So we have the following corollary for the ORA
algorithm:
Corollary 4.1
Under any feasible initial relay node assignment, the ORA algorithm can
find an optimal relay node assignment.
4.8 Numerical examples
In this section, we present some numerical results to illustrate some properties

of the ORA algorithm.
4.8.1 Simulation setting

We consider a 100-node ad hoc network. The location of each node is given
in Table 4.4. For this network, we consider both the cases of Nr Ns and
Nr < Ns . In the first case, we have 30 sourcedestination pairs and 40 relay
nodes. While in the second case, we have 40 sourcedestination pairs and 20
relay nodes. The role of each node (either as a source, destination, or relay)
for each case is shown in Figs. 4.6 and 4.10, respectively, with details given in
Table 4.4.
For the simulations, we assume W = 10 MHz bandwidth for each channel.
The maximum transmission power at each node is set to 1 W. Each relay node
employs AF for CC. We assume that hsd only includes the path-loss com-
ponent between nodes s and d and is given by |hsd |2 = ||s d||4 , where
||s d|| is the distance (in meters) between these two nodes and 4 is the path-
loss index. For the AWGN channel, we assume the variance of noise is 1010
W at all nodes.
4.8.2 Results
Case 1: Nr Ns . In this case (see Fig. 4.6), we have 30 sourcedestination
pairs and 40 relay nodes.
Under ORA, after preprocessing, we start with an initial relay node assign-
ment in the first iteration. Such an initial assignment is not unique. But regard-
less of the initial relay node assignment, we expect the objective value Cmin
to converge to the optimum (Corollary 4.1). To validate this result, we show
83 4.8 Numerical examples
Table 4.4 Locations and roles of all the nodes in the network.
Role Role Role

Location Case 1 Case 2 Location Case 1 Case 2 Location Case 1 Case 2
(75, 500) s1 s1 (220, 190) d4 d4 (380, 370) r7 s31

(170, 430) s2 s2 (660, 190) d5 d5 (300, 350) r8 r8
(170, 500) s3 s3 (430, 630) d6 d6 (410, 650) r9 s33
(250, 650) s4 s4 (180, 620) d7 d7 (470, 500) r10 d40
(400, 550) s5 s5 (750, 625) d8 d8 (660, 525) r11 s39
(340, 230) s6 s6 (310, 480) d9 d9 (600, 425) r12 s40
(390, 150) s7 s7 (1100, 180) d10 d10 (510, 200) r13 s38
(460, 280) s8 s8 (1110, 360) d11 d11 (575, 325) r14 r14
(700, 500) s9 s9 (875, 600) d12 d12 (750, 560) r15 r15
(750, 360) s10 s10 (700, 300) d13 d13 (800, 360) r16 r16
(800, 90) s11 s11 (650, 550) d14 d14 (860, 260) r17 r17
(900, 160) s12 s12 (740, 170) d15 d15 (980, 450) r18 r18
(1125, 300) s13 s13 (410, 810) d16 d16 (950, 310) r19 r19
(1000, 340) s14 s14 (550, 1100) d17 d17 (950, 200) r20 d37
(1025, 540) s15 s15 (150, 790) d18 d18 (100, 1000) r21 s32
(100, 1120) s16 s16 (210, 1110) d19 d19 (310, 980) r22 r12
(150, 920) s17 s17 (530, 720) d20 d20 (250, 800) r23 d32
(330, 1110) s18 s18 (800, 1140) d21 d21 (460, 1010) r24 r13
(450, 890) s19 s19 (1080, 1100) d22 d22 (610, 930) r25 d34
(650, 1050) s20 s20 (940, 790) d23 d23 (680, 760) r26 s34
(700, 640) s21 s21 (1360, 640) d24 d24 (700, 900) r27 r20
(820, 880) s22 s22 (1280, 1120) d25 d25 (910, 1120) r28 d35
(1150, 1060) s23 s23 (1260, 350) d26 d26 (970, 970) r29 s35
(1480, 1120) s24 s24 (1500, 50) d27 d27 (1360, 910) r30 r9
(1160, 720) s25 s25 (1450, 605) d28 d28 (1200, 920) r31 r11
(1050, 50) s26 s26 (1030, 910) d29 d29 (1250, 690) r32 d36
(1350, 450) s27 s27 (1150, 230) d30 d30 (1290, 180) r33 r10
(1380, 110) s28 s28 (80, 370) r1 d31 (150, 360) r34 r5
(1500, 800) s29 s29 (110, 280) r2 r2 (1380, 380) r35 r7
(1500, 300) s30 s30 (160, 300) r3 r3 (1220, 60) r36 s37
(200, 50) d1 d1 (280, 520) r4 r4 (1190, 510) r37 s36
(520, 240) d2 d2 (375, 580) r5 d39 (500, 40) r38 d38
(40, 100) d3 d3 (385, 450) r6 r6 (50, 805) r39 d33
(1510, 920) r40 r1
the results of running the ORA algorithm under two different initial relay node
assignments, denoted as I and II (see Table 4.5).
In Table 4.5, the second column shows the data rate for each source
destination pair under direct transmissions. Note that the minimum rate among
all pairs is 1.83 Mbps, which is associated with s7 . The third to fifth columns
are results under the initial relay node assignment I and the sixth to eighth
columns are results under the initial relay node assignment II. The symbol
denotes direct transmission. Note that initial relay node assignments I and
II are different. As a result, the final assignment is different under I and II.
Figure 4.6 Topology for a (meters)

Senders Receivers Potential relays
100-node network for Case 1 1200
d19 d21 s24
(Nr Ns ), with Ns = 30 and s16 s18 d17
s20
r28 d22
s23
d25
1100
Nr = 40. r21 r24
r22 r29
1000 r40
r25 r27
s17 s19 d29 r31
s22 r30
900
r39 d16
d18 r23 r26 d23 s29
800 s25
r9 d20
s4 d8 r32
d6 s21 d12
700 d7 s5 d14 d24
r15 s15
r5 d28
600 s3 r10 r11 r37
s1 r4 d9 s9
r18
500 r6 s27
r12
s2 s10 r16 d11
r1 s14 r35 r34
400 r7 r14 d26
s8 d13
r3 r8 s13
r17 r19 s30
300 d2
d30
r2 s6
200 r33
d4 r13 d15 r20 d10
d5
s7
s28
100 s12 s26 r36
d3 r38 d27
s11
d1
0
0 100 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500 1600
(meters)
However, the final objective value (i.e., Cmin ) under I and II is identical (4.43
Mbps).
Fig. 4.7 shows the objective value Cmin at each iteration under initial relay
node assignments I and II. Under either initial assignments, Cmin is a nonde-
creasing function of iteration number. Note that a higher initial value of Cmin
does not mean that ORA will converge faster.
Case 2: Nr < Ns . In this case (see Fig. 4.8), we have 40 sourcedestination
pairs and 20 relay nodes.
Table 4.6 shows the results of this case under two different initial relay node
assignments I and II. The second column in Table 4.6 lists the data rate under
direct transmissions for each sourcedestination pair. As discussed at the end
of Section 4.6.2, for the case of Nr < Ns , it is only necessary to consider relay
node assignments for Nr = 20 source nodes corresponding to the 20 small-
est data rates, i.e., nodes s1 , s3 , s4 , s5 , s7 , s8 , s11 , s12 , s13 , s14 , s15 , s16 , s17 , s21 ,
s24 , s25 , s27 , s28 , and s29 .
Again in Table 4.6, the objective value Cmin is identical (3.80 Mbps) regard-
less of different initial relay node assignments (I and II). Note that despite the
different final relay node assignments under I and II, the objective value Cmin
is identical.
Fig. 4.9 shows the objective value Cmin at each iteration under initial relay
node assignments I and II. Again, we observe that in Fig. 4.9, Cmin is a nonde-
creasing function of iteration number under both initial relay node assignments
I and II.
Significance of preprocessing Now we use a set of numerical results to
show the significance of preprocessing in our ORA algorithm. We consider
the same network in Fig. 4.6 with 30 sourcedestination pairs and 40 relay
Table 4.5 Optimal assignments for Case 1 (Nr Ns ) under two different initial relay
node assignments.
Relay assignment I Relay assignment II

CD Final rate Final rate
Session (Mbps) Initial Final (Mbps) Initial Final (Mbps)
s1 2.62 r3 6.54 r3 r3 6.54

s2 4.60 r8 r7 9.46 r8 r7 9.46
s3 3.81 r2 8.73 r1 r1 7.21
s4 2.75 r4 4.66 r4 r4 4.66
s5 3.15 r14 6.47 r7 r14 6.47
s6 4.17 r6 9.25 r10 r6 9.25
s7 1.83 r6 r8 4.76 r6 r8 4.76
s8 2.99 r12 7.22 r16 r12 7.22
s9 4.92 r12 r10 9.81 r12 r10 9.81
s10 4.80 r18 4.80 4.80
s11 4.13 r16 r20 9.13 r17 r20 9.13
s12 3.23 r19 5.89 r18 r18 5.55
s13 3.68 r18 4.84 r19 r17 7.32
s14 4.23 r16 7.87 r15 r15 5.29
s15 2.62 r17 r17 4.86 r20 r19 5.84
s16 3.30 r22 7.29 r22 r22 7.29
s17 4.17 r24 5.62 r24 r24 5.62
s18 6.03 r21 r21 7.37 r23 r23 6.26
s19 8.76 8.76 8.76
s20 6.95 6.95 6.95
s21 1.90 r27 r27 4.90 r27 r27 4.90
s22 7.65 r28 r28 8.71 r28 r28 8.71
s23 7.55 r29 r29 11.26 r29 r28 11.26
s24 2.12 r40 r40 4.43 r40 4.43
s25 3.90 r30 5.87 r30 5.87
s26 6.08 r36 r36 6.81 r36 r36 6.81
s27 3.61 r34 5.44 r34 5.44
s28 2.04 r35 r35 5.29 r35 5.29
s29 2.32 r30 r31 4.68 r31 4.68
s30 6.60 r34 r33 9.65 r34 r33 9.65
nodes, but we remove the preprocessing step in the ORA algorithm. As an

example, the third column of Table 4.7 shows an initial assignment without
first going through the preprocessing step. Although the objective value Cmin
also reaches the same optimal value (4.43 Mbps) as that in Table 4.5, the final
data rate for some nonbottleneck source nodes could be worse than under direct
transmissions. For example, for s19 , its final rate is 4.81 Mbps, which is less
than the direct transmission rate (8.76 Mbps). Such an event is undetectable
without the preprocessing step, as 4.81 Mbps is still greater than the optimal
objective value (4.43 Mbps).
Figure 4.7 Case 1 (Nr Ns ):

The objective value Cmin at
each iteration of the ORA
algorithm under two different
initial relay node assignments.
Figure 4.8 Topology for a (meters)

Senders Receivers Potential relays
100-node network for Case 2 1200
d19 d21 s24
(Nr < Ns ), with Ns = 40 and 1100
s16 s18 d17
s20
d35 d22
s23
d25
Nr = 20. s32 r13

r12 s35
1000 r1
d34 r20
s17 s19 d29 r11
s22 r9
900
d33 d16
d18 d32 s34 d23 s29
800
s33 d20 s25
s4 d8 d36
d7 d6
700 s21 d12
s5 d14 d24
d39 r15 s15
d28
600 s1 s3 d40 s36
r4 s39
d9 s9
s2 r18
500 r6 s40 s27
s31 s10 r16 d11
s14 r7 r5
d31 r14 d26
400 s8
r3 d13
r8 s13
r17 r19 s30
300 d2
s6 d30
r2 d4 s38 d37
d5 d15
200 s7 r10
d10
d3 s11 s28
100 d1 s12 s26 s37
d38 d27
0
0 100 200 300 400 500 600 700 800 900 1000 1100 12001300 1400 1500 1600
(meters)
On the other hand, when the preprocessing step is employed, ORA can
ensure that the final rate for each sourcedestination pair is no less than that
under direct transmission, as shown in Table 4.5.
4.9 Chapter summary
This chapter illustrates how an optimization problem can be solved follow-

ing algorithmic techniques from computer science (CS). For certain prob-
lems, general optimization methods from operations research (OR) may not
always be the best approach. In fact, a formulation following ORs optimiza-
Table 4.6 Optimal assignments for Case 2 (Nr Ns ) under two different initial relay
node assignments.
Relay assignment I Relay assignment II

CD Final rate Final rate
Session (Mbps) Initial Final (Mbps) Initial Final (Mbps)
s1 2.62 r2 6.62 r3 r3 6.54

s2 2.60 4.60 4.60
s3 3.81 3.81 3.81
s4 2.75 r5 4.66 r8 r5 5.20
s5 3.15 r6 3.80 r14 r6 3.80
s6 4.17 4.17 4.16
s7 1.83 r8 4.76 r6 r8 4.76
s8 2.99 r14 4.43 r14 4.43
s9 4.92 4.92 4.92
s10 4.80 4.80 4.80
s11 4.13 4.13 4.13
s12 3.23 r18 5.55 r18 5.55
s13 3.68 r19 8.04 r19 8.04
s14 4.23 4.23 4.23
s15 2.62 r16 5.60 r16 5.60
s16 3.30 r12 7.30 r12 7.30
s17 4.17 4.17 4.17
s18 6.03 6.03 r13 6.03
s19 8.76 8.76 r12 r13 8.97
s20 6.95 6.95 r20 6.95
s21 1.90 r20 4.90 r20 4.90
s22 7.65 7.65 7.65
s23 7.55 7.55 r11 7.55
s24 2.12 r9 5.15 r9 5.15
s25 3.91 3.91 3.91
s26 6.08 6.08 r10 6.08
s27 3.61 r10 5.27 r10 5.27
s28 2.04 r7 5.29 r7 5.29
s29 2.32 r11 4.68 r11 4.68
s30 6.60 6.60 6.60
s31 11.06 11.06 11.06
s32 17.47 17.47 17.47
s33 4.86 4.86 4.86
s34 31.34 31.34 31.34
s35 37.87 37.87 37.87
s36 29.79 29.79 29.79
s37 10.65 10.65 10.65
s38 38.27 38.27 38.27
s39 12.10 12.10 12.10
s40 41.70 41.70 41.70
Figure 4.9 Case 2 (Nr < Ns ):

The objective value Cmin at
each iteration of the ORA
algorithm under two different
initial node assignments.
Table 4.7 An example illustrating the importance of preprocessing.
Without preprocessing
Sender CD (Mbps) Initial Final Final rate (Mbps)
s1 2.62 r3 r3 6.54
s2 4.60 4.60
s3 3.81 r2 8.73
s4 2.75 r8 r4 4.66
s5 3.15 r14 r14 6.47
s6 4.17 r6 9.25
s7 1.83 r6 r8 4.76
s8 2.99 r12 7.22
s9 4.92 4.92
s10 4.80 4.80
s11 4.13 r20 9.13
s12 3.24 r18 5.55
s13 3.68 r17 7.32
s14 4.23 r16 7.87
s15 2.62 r19 5.84
s16 3.30 r22 7.30
s17 4.17 r24 5.62
s18 6.03 r23 r21 7.37
s19 8.76 r39 r39 4.81
s20 6.95 r26 r26 7.25
s21 1.90 r27 4.90
s22 7.65 r28 r28 8.71
s23 7.55 r29 r29 11.26
s24 2.12 r40 4.43
s25 3.91 r30 5.87
s26 6.08 r33 r33 7.55
s27 3.61 r34 5.45
s28 2.04 r35 5.29
s29 2.33 r31 4.68
s30 6.60 6.60
89 4.10 Problems
tion approach may lead to a solution with non-polynomial-time complexity.

But a solution with nonpolynomial-time complexity does not mean that the
problem is not in P. In fact, we may well develop a different algorithm to solve
the problem with polynomial-time complexity. This is what we have illustrated
in this chapter, where we developed a polynomial-time algorithm.
In our case study, we considered a relay node assignment problem in CC.
Our objective is to assign a set of available relay nodes to different source
destination pairs so as to maximize the minimum data rate among all the pairs.
Following the OR optimization approach, we showed that the problem can
be formulated as an MILP, which is NP-hard in general. But this does not
mean that the problem is NP-hard. Instead, by following a CS algorithm design
approach, we developed a polynomial-time exact algorithm for this problem.
A novel idea in this algorithm is a linear marking mechanism, which is able to
achieve polynomial-time complexity at each iteration. We gave a formal proof
of the optimality of the algorithm.
This chapter concludes the first part of this book, which is on developing
optimal solutions.
4.10 Problems
4.1 Can a generally NP-hard formulation for a problem be used to claim that
the problem is NP-hard? Why?
4.2 Verify Eq. (4.6) in this chapter.
4.3 Show that IAF (SNRsd , SNRsr , SNRrd ) is an increasing function of Ps
and Pr , respectively. Do the same for IDF (SNRsd , SNRsr , SNRrd ).
4.4 Consider the simple three-node relay channel in Fig. 4.1. The locations
of the source node s and the destination node d are at (0, 0) and (100, 0),
respectively, all in meters. The maximum transmit power of both source and
relay nodes is 0.1 W. The channel coefficient between two nodes is h = 2d 2 ,
where d is the distance between the two nodes. The noise power spectrum
density (PSD) at all nodes is 2 1016 W/Hz and the channel bandwidth is 10
MHz.
(a) Calculate the achievable rate under direct transmission between s and d.
(b) If the location of the relay node r is at (50, 33) (in meters), calculate the
achievable rates under AF and DF, respectively.
(c) If the location of the relay node r is at (80, 50) (in meter), calculate the
achievable rates under AF and DF, respectively.
4.5 If DF is employed for CC, how will it affect the operation of ORA?
Justify your answer.
Table 4.8 Initial relay node assignment for a CC net-

work with four sessions and six relay nodes.
r1 r2 r3 r4 r5 r6
s1 4 17 8 15 5 8 15
s2 6 12 8 16 14 8 13
s3 3 15 6 21 13 14 13
s4 11 19 15 12 16 9 15
Table 4.9 Initial relay node assignment for a CC net-

work with seven sessions and seven relay nodes.
r1 r2 r3 r4 r5 r6 r7
s1 1 17 6 15 10 20 11 15
s2 6 11 6 17 9 18 10 13
s3 6 5 18 20 16 18 15 12
s4 4 18 12 9 6 15 19 11
s5 15 19 4 15 16 19 8 15
s6 9 6 11 16 15 18 11 15
s7 6 11 19 5 7 9 18 15
4.6 In practice, the number of relay nodes may or may not be less than the
number of source nodes, i.e., Nr < Ns or Nr Ns . How will this affect the
operation of the ORA algorithm?
4.7 Describe the marking mechanism in the ORA algorithm. Why do we call
it linear? Explain its role in the algorithms complexity.
4.8 For the naive marking mechanism described in Section 4.6.3, show that
the number of checks for relay nodes within an iteration is O(Nr Nr !) if Nr <
r Nr !
Ns and O( (NrNN s +1)!
) if Nr Ns .
4.9 For Example 4.1 in Section 4.6.2, give the details of how ORA and linear
marking work in the second and third iterations, respectively.
4.10 Apply the ORA algorithm to find an optimal solution for the data given
in Table 4.8. Show your work for each iteration.
4.11 Apply the ORA algorithm to find an optimal solution for the data given
in Table 4.9. Show your work for each iteration.
4.12 Give a complete proof of Claim 4.2.
4.13 For any given relay node assignment with an objective value Cmin , we
can define a directional graph as follows (see Fig. 4.10 as an example). The set

of nodes in the graph is Ns Nr {S, D, }, where S, D, and are dummy
source, destination, and relay nodes, respectively. The graph has the following
91 4.10 Problems
Figure 4.10 A sample for the r1

1
construct graph in Problem 4.13.
s1
1
1
1 1 r2
s2
1
1
1
S
D
1
1
1 1
rNr Ns
sNs
1
edges: (i) edges (S, si ) with unit capacity, for each si Ns ; (ii) edges (si , rj )
with unit capacity, for each si Ns , rj Nr , CR (si , rj ) > Cmin , and edges
(si , ) with unit capacity, for each si Ns , CR (si , ) > Cmin ; and (iii) edges
(rj , D) with unit capacity, for each rj Nr , and an edge (, D) with capacity
Ns . Show how such a max-flow problem for the defined graph is related to the
problem of identifying an improving solution.
4.14 In each iteration of the ORA algorithm, we try to find a better solution.
Denote C as the set of all possible values for the maximum Cmin value in an
optimal solution. Then in the worst case, ORA decreases |C| by one in an itera-
tion. We can improve the ORA algorithm as follows. In each iteration, instead
of just finding a better solution, we identify the median value Cmed for the cur-
rent set C and try to find a solution with an objective value of at least Cmed . We
may or may not find such a solution. But for either case, we can decrease |C|
to 12 |C| and thus the number of iterations can be significantly decreased. Give
a detailed description of this enhanced ORA (EORA) algorithm and analyze
its complexity. (Hint: The complexity of one iteration in EORA is higher than
that for ORA, since it may be necessary to find better relay nodes for multiple
sources in an iteration of EORA.)
PART
II Methods for Near-optimal and

Approximation Solutions
CHAPTER
5 Branch-and-bound framework
and application
Vision is not enough, it must be combined with venture. It is not enough to stare
up the steps, we must step up the stairs.
Vaclav Havel
5.1 Review of branch-and-bound framework
In this chapter, we present a general and effective approach to solve NP-hard

or NP-complete problems, despite their potential exponential worst-case com-
plexity. Although the complexity of this approach is nonpolynomial, when
designed appropriately this approach may still be a viable approach to solve
many problems that arise in practice, particularly when their problem sizes are
not terribly large.
The approach that we describe is called branch-and-bound [113]. It is a
powerful general purpose framework to solve nonconvex programming prob-
lems. Such a framework aims to obtain a (1 )-optimal solution for a small
given 0. Obviously, the smaller is, the higher is the complexity. Thus, we
need to select a suitable based on the optimality requirement. The efficiency
of the algorithm depends on its ability to correctly remove large portions of
the solution space during each iterative step within the process of identifying
the best solution. As a result, we progressively focus on smaller and smaller
portions of the solution space and thus we are able to find a (1 )-optimal
solution much faster than a brute-force exhaustive search.
We now describe the branch-and-bound framework for a minimization prob-
lem (the procedure for a maximization problem is similar). Initially, we need
to determine the solution space for our problem, including the set of feasible
values for each of the optimization variables. The framework consists of two
steps, namely, the bounding step and the branching step.
95
96 Branch-and-bound framework and application
Bounding step The goal of this step is to develop a lower bound and an
upper bound for a particular problem. First of all, by using some relaxation
technique, branch-and-bound obtains a relaxation for the original problem and
its solution provides a lower bound (LB) for the objective function. Such a
relaxation is usually in the form of a LP or a convex programming problem,
which we can solve in polynomial time (given that the relaxation itself is of
polynomial size). It is clear that we should make this relaxation as tight as
possible, i.e., the obtained lower bound should be as close to the optimal value
as possible. The particular approach to develop a tight and polynomial-time
solvable relaxation needs to be carefully designed by the user.
Using the relaxation solution as a starting point, branch-and-bound employs
a local search algorithm to find a feasible solution to the original problem,
which provides an upper bound (U B) for the objective function (see Fig. 5.1(a)
for an example). Again, the local search algorithm needs to be carefully
designed by the user. This computation of lower and upper bounds constitutes
the bounding step of the branch-and-bound framework.
If the obtained lower and upper bounds are within an -tolerance of each
other, i.e., LB (1 )U B, we are done, i.e., the current feasible solution is
(1 )-optimal. If the relaxation is not tight, then the lower bound LB could
be significantly lower than the upper bound U B. To close this gap, we must
attempt to produce a tighter relaxation that yields a smaller relaxation error.
This can be achieved by a branching (or called a partitioning) step.
Branching step The branching step includes partitioning problem, updating
lower and upper bounds, and/or fathoming a subproblem.
Partitioning problem To have a small relaxation error, we can narrow down

the value-sets of variables that affect this error. We call these variables par-
titioning variables. In general, branch-and-bound selects a partitioning vari-
able that evidently contributes most toward the relaxation error and splits its
value-set into two (or sometimes more) sets based on its value in the relax-
ation solution (see Fig. 5.1(b)). For example, if a variable x with a value-set
{0, 1, . . . , 10} is selected and its value in the relaxation solution is 3.7, then
its value-set can be partitioned into two sets {0, 1, 2, 3} (the set of values
no more than 3.7) and {4, 5, . . . , 10} (the set of values more than 3.7). Note
that if xs value in the relaxation solution is very close to lower bound xL or
upper bound xU , i.e., min{x xL , xU x} < 0.1(xU xL ), then we should
use the bisection rule. That is, if xs value in the relaxation solution is 0.7
(we have min{0.7 0, 10 0.7} < 0.1(10 0)), then its value-set can be
partitioned into two sets {0, 1, . . . , 4} and {5, 6, . . . , 10}. This is called par-
titioning or branching (dichotomous branching in this case, where an origi-
nal value-set is split into two, which we assume to be the case in the sequel).
Then the original problem (denoted as Problem 1) is divided into two new
problems (denoted as Problem 2 and Problem 3).
97 5.1 Review of branch-and-bound framework
Figure 5.1 Illustration of the

branch-and-bound framework.
UB=UB1
Objective
LB=LB1
Original problem 1
(a) Iteration 1.
UB3
UB=UB2
Objective
LB2
LB=LB3
Problem 2 Problem 3
(b) Iteration 2.
UB2
UB4
Objective
UB=UB5
LB4 LB5
LB=LB2
Problem 2 Problem 4 Problem 5
(c) Iteration 3.
Algorithm 5.1 The branch-and-bound framework for a minimization problem

Branch-and-bound framework
Initialization
1. Let the initial upper bound U B = and the initial problem list
include only the original problem, denoted as Problem 1.
2. Determine the initial value-set for each partitioning variable.
3. Build a linear relaxation for Problem 1 and obtain its solution 1 .
4. The objective value of 1 is a lower bound LB1 to Problem 1.
Main iteration
5. Select Problem z that has the minimum LBz -value among all
problems in the problem list.
6. Update the lower bound by setting LB = LBz .
7. Find a feasible solution z starting from z via a local search
algorithm and denote its objective value as U Bz .
8. If (U Bz < U B) {
9. Update = z and U B = U Bz .
10. If LB (1 )U B, stop with the (1 )-optimal solution .
Otherwise, remove all Problems z with LBz (1 )U B
from the problem list. }
11. Select a variable that appears to contribute most toward the relaxation
error and partition its value-set into two sets based on its value in z .
12. Build two new Problems z1 and z2 based on the two sets in Step 11.
13. Remove Problem z from the problem list.
14. Obtain LBz1 and LBz2 for Problems z1 and z2 via their relaxations.
15. If LBz1 < (1 )U B, add Problem z1 into the problem list.
16. If LBz2 < (1 )U B, add Problem z2 into the problem list.
17. If the problem list is empty, stop with the (1 )-optimal solution
. Otherwise, go to Step 5.
Updating lower and upper bounds For each of two new problems, the
branch-and-bound procedure solves a corresponding relaxation and performs
a local search. Hence, we derive lower bounds LB2 and LB3 for Problems
2 and 3, respectively. We also possibly obtain feasible solutions that provide
upper bounds U B2 and U B3 for Problems 2 and 3, respectively. It is possible
that a new problem does not have a feasible solution or the (heuristic) local
search procedure might fail to find a feasible solution even if one exists. But
since the feasible solution 1 to Problem 1, assuming that such a solution has
been detected, must be feasible to one of the two new problems (unless some
explored subregion that contains 1 has been deleted from consideration
during the analysis of Problem 1), we have at hand a feasible solution for at
least one new problem by using the previous feasible solution if necessary.
Since the relaxations in Problems 2 and 3 are both more tightly constrained
99 5.1 Review of branch-and-bound framework
than that of Problem 1, we have min{LB2 , LB3 } LB1 . The lower bound of
the original problem is updated from LB = LB1 to LB = min{LB2 , LB3 }.
The upper bound of the original problem is updated from U B = U B1 to
U B = min{U B, U B2 , U B3 }, since the best feasible solution to the original
problem is the solution with the smallest U Bi -value.1 As a result, we now
have a smaller gap between LB and U B. If LB (1 )U B, we are done.
Otherwise, we choose a problem with the minimum lower bound (Problem
3 in Fig. 5.1(b)) and perform partitioning on this problem.
Fathoming a subproblem. Note that during the branch-and-bound process,
if we find a Problem z with LBz (1 )U B (see Problem 4 in Fig. 5.1(c)),
then we have the following two possible cases:
Case 1: The global optimal solution is not in Problem z. In this case, the
removal of Problem z will not cause the loss of the optimal solution in future
iterations.
Case 2: The global optimal solution is in Problem z. In this case, the opti-
mal (feasible) solution must have an objective value greater than or equal to
LBz , which means that it is also greater than or equal to (1 )U B (since
LBz (1 )U B). Thus, the current best feasible solution with objective
value U B is already a (1 )-optimal solution. Therefore, we can still guar-
antee (1 )-optimality if we remove this problem from the list for further
consideration.
The foregoing cases underscore the principal advantage of the branch-and-
bound framework, i.e., we can remove a problem before we completely solve
it. This removal of a subproblem from consideration in the active list of
problems is called fathoming the subproblem. Eventually, when we obtain
LB (1 )U B, the branch-and-bound procedure terminates.
It has been shown that, under very general conditions, a branch-and-bound

solution procedure always converges [146]. Moreover, although the worst-
case complexity of such a procedure is exponential, the actual running time
could be fast when all partitioning variables are reasonably bounded inte-
gers (e.g., in the case study problem considered in this chapter as well as
many other problems in wireless networks), and the relaxations derived are
tight.
Algorithm 5.1 presents a pseudocode for the branch-and-bound framework.
As discussed, there are several key components in this framework that are
problem specific and need to be carefully designed. In this chapter, we will
demonstrate how to design these key components in a case study.
1 Note that we always keep the best detected (incumbent) feasible solution in the process.
5.2 Case study: Power control problem for multi-hop

cognitive radio networks
The case study that we present in this chapter is on a power control prob-
lem for a multi-hop cognitive radio network (CRN). We will show how a
branch-and-bound framework can be applied to solve this problem. Cognitive
radio (CR) is an enabling physical layer technology to enhance the efficiency
of radio spectrum [173]. A CR is a frequency-agile data communication device
that has a rich control and monitoring (spectrum sensing) interface [65; 112].
It capitalizes advances in signal processing and radio technology, as well as
recent advancements in spectrum policy [130]. A frequency-agile radio mod-
ule is capable of sensing the available bands [45; 47; 111; 160], reconfiguring
RF, and switching to newly selected frequency bands. Thus, a CR can be pro-
grammed to tune and operate on specific frequency bands over a wide spectrum
range [130]. From an application perspective, CR allows a single radio to pro-
vide a wide variety of functions, acting as a cell phone, broadcast receiver,
GPS receiver, wireless data terminal, etc.
A fundamental problem for a wireless network is power control. The benefits
of per-node-based power control for CRNs were discussed in [150]. In a multi-
hop wireless network, power control is challenging since it directly affects
the upper layers (scheduling and routing). When each node is allowed to per-
form power control (which we call per-node-based power control), the problem
becomes even more difficult due to its large optimization space. Due to such
difficulty, some related work, e.g., [13; 86], assumed synchronized power con-
trol, where transmission power at each node in the network is adjustable but
is synchronized to be identical. Needless to say, such simplification can hardly
reflect real problems in practice.
The main difficulty associated with the per-node-based power control prob-
lem is its tight coupling to scheduling and routing. Therefore, a joint formu-
lation of multiple layers is needed, which inevitably leads to a very complex
problem. Although there has been some success in the context of asymptotic
scaling laws (e.g., [60; 86]), theoretically optimal results for a given finite-sized
network remain limited. For example, in [16], Bhatia and Kodialam optimized
power control and routing, but assumed some frequency hopping mechanism
was in place for scheduling, which helped simplify the joint consideration
of scheduling. In [37], Elbatt and Ephremides optimized power control and
scheduling, but assumed routing was given a priori. Although [26; 32] aimed
to investigate joint power control, scheduling, and routing problems, both of
the proposed procedures followed a de-coupled approach, where the final
solution was obtained one layer at a time (instead of solving a joint optimiza-
tion problem, as we will do in this chapter). Due to such de-coupling, the final
solution can only be suboptimal at best.
101 5.3 Mathematical modeling
In this chapter, we study the per-node-based power control problem for a

multi-hop CRN. This problem is both challenging and interesting as it inherits
not only all the difficulties associates with per-node-based power control, but
also all the unique characteristics associated with a CRN. We develop a for-
mal mathematical model for a joint per-node-based power control, scheduling,
and flow routing problem. This joint formulation leads to a mixed-integer non-
linear programming (MINLP) problem. Subsequently, we develop a unified
(instead of a de-coupled) solution procedure based on the branch-and-bound
framework and a convex hull relaxation. This solution procedure guarantees a
(1 )-optimal solution, where 0 is a small prespecified error tolerance
parameter.
The remainder of this chapter is organized as follows. In Section 5.3,
we develop a unified mathematical model for per-node-based power control,
scheduling, and flow routing. In Section 5.4, we formulate the cross-layer
optimization problem. Section 5.5 applies the branch-and-bound framework
to design a solution procedure. In Section 5.6, we use numerical results to
illustrate the solution procedure. Section 5.7 summarizes this chapter.
5.3 Mathematical modeling
Consider a multi-hop CRN consisting of a set of N nodes. The set of available

frequency bands at each node depends on its location and may not be the same.
For example, at node i, its available frequency bands may consist of bands
I, III, and V, while at a different node j , its available frequency bands may
consist of bands I, IV, and VI, and so forth. More formally, denote by Mi the
set of available frequency bands at node i. For simplicity, we assume that the
bandwidth of each frequency band is W . Denote M as the set of all frequency

bands present in the network, i.e., M = iN Mi . Table 5.1 lists the notation
used in this chapter.
5.3.1 Necessary and sufficient condition for successful transmission

Scheduling for transmission at each node in the network can be done either
in-time or frequency domain. In this chapter, we consider scheduling in the
frequency domain in the form of frequency bands. Similar modeling can also
be done in-time domain following the same token.
Consider a transmission from node i to node j . Suppose that band m is
(
available at both node i and node j , i.e., m Mij , where Mij = Mi Mj .
Denote pij m as the transmission power from node i to node j in frequency band
m. For transmission from node i to node j , a simple model for path attenuation
loss gij is
Table 5.1 Notation.
Symbol Definition
m
cij Link capacity of link i j under pij m
d(l) Destination node of session l L

dij Distance between nodes i and j
fij (l) Data rate that is attributed to session l on link i j
f The vector of variables fij (l), l L, i N , i = d(l), j Ti , j = s(l)
gij Path attenuation loss from node i to node j
Ijm The set of nodes that can interfere with node j on band m (under
full transmission power P )
L The set of user communication sessions
Mi Theset of available bands at node i N
M = iN ( Mi , the set of bands in the network
Mij = Mi Mj , the set of available bands for link i j
m
pij Node is transmission power to node j on band m
P The maximum transmission power at a transmitter
PT The threshold of minimum receiving power that can be decoded
PI The threshold of maximum interference power that is negligible
m
qij Transmission power level from node i to node j on band m
q The vector of variables qijm, i N , m M , j T m
i i
Q Number of transmission power levels
r(l) Rate requirement of session l
RT (p), RI (p) The transmission and interference ranges under power p
RTmax , RImax The maximum transmission and interference ranges under full
transmission power P
s(l) Source nodes of session l L
Tim The set of nodes to which node i can transmit on band m (under
fulltransmission power P )
Ti = mMi Tim , the set of nodes to which node i can transmit on
all bands (under full transmission power P )
W Bandwidth of each frequency band
m
xij A binary indicator to denote whether or not band m is used by
link i j
x The vector of variables xijm, i N , m M , j T m
i i
Path-loss index
Ambient Gaussian noise density
Notation for the branch-and-bound framework

LBz , LB The lower bounds for Problem z and the original problem
U Bz , U B The upper bounds for Problem z and the original problem
The desired accuracy in the final solution
z The set of all possible values of (x, q) in Problem z
z The solution obtained by local search for Problem z
A (1 )-optimal solution
gij = dij , (5.1)
where dij is the physical distance between nodes i and j and is the path-
loss index. In this context, we assume data transmission from node i to node
j is successful only if the received power at node j exceeds a threshold PT .
Denote the transmission range of node i under pijm as R (p m ). Then, based on
T ij
gij pij PT and (5.1), we can calculate the transmission range of this node
m
as follows:
# m
$1/
pij
m
RT (pij ) = . (5.2)
PT
Since the receiving node j must be physically within the transmission range of
node i, we have
# m $1/
pij
(C-1) dij .
PT
Similarly, we assume that an interference is nonnegligible only if it exceeds

a threshold, say PI , at a receiver. Denote the interference range of node k
(k N , k = i) under pkh
m as R (p m ), where h is the intended receiving node
I kh
of transmitting node k. Then following the same token as the derivation for
the transmission range, we can obtain the interference range of node k as
m 1/
m ) = pkh
RI (pkh . Since the receiving node j must not fall in the interfer-
PI
ence range of any other node k that is transmitting in the same band, we have
% m &1/
pkh
(C-2) dj k .
PI
As an example, Fig. 5.2(a) shows a network with three links (1 2, 3 4,
and 5 6). For each transmitting node (1, 3, and 5), the inner circle (dashed)
represents the transmission range and the outer circle (solid) represents the
interference range. Clearly, each receiving node falls in the transmission range
of its respective transmitting node. Further, we can see that both receiving
nodes 2 and 6 fall in the interference range of node 3. Thus, when link 3 4
is using a frequency band m for transmission, links 1 2 and 5 6 should
not use the same band. It should also be noted that when link 3 4 is not
using a frequency band m, both links 1 2 and 5 6 may use band m. This
is because receiving node 2 does not fall in node 5s interference range and
receiving node 6 does not fall in node 1s interference range. Now, consider
that each node can adjust its transmission power. In this setting, nodes 1, 3,
and 5 can reduce their transmission powers while still maintaining data trans-
mission to the corresponding receiving nodes (see Fig. 5.2(b)). Then receiving
nodes 2 and 6 are no longer in node 3s interference range. As a result, both
transmitting nodes 1 and 5 can also transmit on band m as node 3 does.
Figure 5.2 A 3-link network.
4
3
6
2 5
(a) No power control case.
4
3
6
5
2
(b) Per-node-based power control case.
5.3.2 Per-node-based power control and scheduling

In this section, we formulate a mathematical model for the joint relationship
between per-node-based power control and scheduling. Suppose m Mij .
Denote
'
1 if node i transmits data to node j on frequency band m,
xijm =
0 otherwise.
As mentioned earlier, we consider scheduling in the frequency domain and
assume unicast communication. Thus once a band m is used by node i for
transmission to node j , this band cannot be used again by node i to transmit to
a different node. That is,

(C-3) xijm 1,
j Tim
where Tim is the set of nodes to which node i can transmit on band m under
full power P .
Denote by RTmax the maximum transmission range of a node when it
transmits at full power P . Then based on (5.2), we have RTmax = RT (P ) =
1/
P
PT . Thus, we have PT = (R max P
. Then, for a node transmitting at a
T )
power p [0, P ], its transmission range is given by
% &1 1 p 1
p p(RTmax )

RT (p) = = = RTmax . (5.3)
PT P P
Similarly, denote RImax as the maximum interference range of a node when it

transmits at full power P . Then, following the same token, we have RImax =
1/
RI (P ) = PPI and PI = (R max
P
)
. For a node transmitting at a power p
I
[0, P ], its interference range is given by
p 1

RI (p) = RImax . (5.4)
P
Recall that Tim denotes the set of nodes to which node i can transmit on
band m under full power P . More formally, we have
Tim = {j : dij RTmax , j = i, m Mj }.
Similarly, denote by Ijm the set of nodes that can interfere with node j on band
m under full power P . Then,
Ijm = {k : dj k RImax , m Mk }.
Note that the definitions of Tim and Ijm are both based on full transmission
power P . When the power level p is below P , the corresponding transmission
and interference ranges will be smaller. Nevertheless, the set of nodes that fall
in the transmission range and the set of nodes that can produce interference
can be upper bounded by those sets under full transmission power.
By (C-1) and (C-2) (i.e., constraints for successful transmission from node
i to node j ), (5.3) and (5.4), we have
# m
$1
pij
dij RT (pij
m
)= RTmax ,
P
% m & 1
pkh
dj k m
RI (pkh ) = RImax (k Ijm , k = i, h Tkm ).
P
Based on the above two constraints, we have the following requirements for
the transmission link i j and the interfering link k h:
) ! d "
m if xijm = 1,
ij
R max P, P
pij T
=0 if xijm = 0.
) d
kj
RImax
P if xijm = 1,
pkh
m
(k Ijm , k = i, h Tkm ).
P if xijm = 0.
Mathematically, these requirements can be rewritten as follows:
% &
dij
(C 1 ) pij
m
P xij , P xij ,
m m
RTmax
% &
dkj
(C 2 ) pkh P 1
m
P xijm (k Ijm , k = i, h Tkm ).
RImax
In addition, for a successful scheduling in the frequency domain, the follow-
ing two constraints must also hold:
(C-4) For a band m Mj that is available at node j , this band cannot be

used for both transmission and receiving. That is, if band m is used at node
j for transmission (or receiving), then it cannot be used for receiving (or
transmission).
(C-5) Similar to constraint (C-3) on transmission, node j cannot use the same
band m Mj to receive from two different nodes.
Note that (C-4) can be viewed as a self-interference avoidance constraint

where, at the same node j , its transmission to another node h on band m inter-
feres its reception from node i on the same band. It turns out that the above
two constraints are mathematically embedded in (C-1 ) and (C-2 ). That is,
once (C-1 ) and (C-2 ) are satisfied, then both the constraints (C-4) and (C-5)
are also satisfied. This result is formally stated in the following lemma:
Lemma 5.1
If transmission powers on every transmission link and interference link sat-
isfy (C-1 ) and (C-2 ) in the network, then (C-4) and (C-5) are also satisfied.
Proof. We first prove that (C-1 ) and (C-2 ) lead to (C-4). To do this, we let
k = j in (C-2 ). Then (C-2 ) degenerates to pjmh P P xijm since djj = 0.
Suppose that node j is receiving from node i on band m, i.e., xijm = 1. Then,

dj h
pjmh P P xijm = 0. Since pjmh R max P xjmh from (C-1 ), we have that
T
xjmh must be 0. That is, if node j is receiving from node i on band m, then
node j cannot transmit to node h in the same band.
Now, suppose that node j is transmitting to node h on band m, i.e., xjmh = 1.

We will show that this implies xijm = 0. We show this by contradiction. That
is, if xijm = 1, then we have just proved in the foregoing case that xjmh = 0.
But this contradicts our initial assumption that xjmh = 1. Therefore, xijm must
be 0. That is, if node j is transmitting to node h on band m, then node j
cannot use the same band for receiving from a node i.
Combining the above two results, we have shown that (C-4) holds.
We now prove that (C-1 ) and (C-2 ) also imply (C-5). Again, the proof is
based on contradiction. Suppose that (C-5) does not hold. Then node j can
receive from two different nodes i and k on the same band m, i.e., xijm = 1 and
m = 1. Note that link k j can be viewed as an interfering link with respect
xkj
to link i j . This corresponds
to letting
h = j in (C-2 ). Then from (C-2 ),
dkj
since xijm = 1, we have pkj
m
RImax
P . Now consider node k is transmitting

dkj
to node j . Then by (C-1 ), we have pkj
m
RTmax P . However, the above two
inequalities cannot hold at the same time since we have RImax > RTmax . This
gives us a contradiction. Thus, the initial assumption that (C-5) does not hold
is incorrect.
The significance of Lemma 5.1 is that, since (C-4) and (C-5) are embedded
in (C-1 ) and (C-2 ), it is sufficient to consider only (C-1 ), (C-2 ), and (C-3) in
the problem formulation. This helps reduce the number of constraints.
5.3.3 Flow routing and link capacity constraints

We assume that there is a set of L active user communication (unicast) sessions
in the CRN. Denote s(l) and d(l) as respectively the source and destination
nodes of session l L, and r(l) as the rate requirement (in b/s) of session l.
Again, allow flow splitting between a source node and its destination node.
The benefits of flow splitting were discussed in Chapter 2.
Mathematically, this can be easily modeled based on flow balance at each
node. Denote by fij (l) the data rate on link (i, j ) that is attributed to session l,

where i N , i = d(l), j Ti = mMi Tim , j = s(l). Note that for fij (l),
we set i = d(l) to ensure that the destination node d(l) will be a sink node and
will not transmit data to a relay node i. We also set j = s(l) to ensure that a
relay node i will not transmit data back to the source node s(l). If node i is the
source node of session l, i.e., i = s(l), then

fij (l) = r(l). (5.5)
j Ti
If node i is an intermediate relay node for session l, i.e., i = s(l) and i =

d(l), then
=s(l)
j k=
d(l)
fij (l) = fki (l). (5.6)
j Ti kTi
If node i is the destination node of session l, i.e., i = d(l), then

fki (l) = r(l). (5.7)
kTi
It can be easily verified that once (5.5) and (5.6) are satisfied, (5.7) must also
be satisfied. As a result, it is sufficient to just include (5.5) and (5.6) in the
formulation.
m as link capacity of link i j under p m . In addition to the above
Denote cij ij
flow balance equations at each node i N for session l L, the aggregated
flow rates on each radio link cannot exceed this links capacity. We have
s(l)=
j,d(l)=i % &
gij m
fij (l) m
cij = W log2 1+ p , (5.8)
W ij
lL mMij mMij
where is the ambient Gaussian noise density.
5.4 Problem formulation
Objective function In this chapter, we consider how to minimize network

resource usage to support a set of user sessions. Network resource usage can
be defined in a number of ways, which typically includes bandwidth usage.
However, bandwidth usage can only quantify resource usage in spectrum, but
cannot take into account of the impact (i.e., interference) of radio transmission
in space. For example, a node transmitting with the same channel bandwidth
but with different power levels will produce different interference footprint
areas. To account for a CRs impact on both bandwidth usage in spectrum
and interference footprint in space, we introduce the so-called bandwidth-
footprint-product (BFP) metric in this chapter (also known as space-bandwidth
product in [96]).
Since each node in the network will use a number of bands for transmission
and each band will have a certain footprint corresponding to its transmission
power, our objective is to minimize network-wide BFP, which is the sum of
BFPs among all the nodes in the network, i.e.,

min W (RI (pij m 2
)) . (5.9)
iN mMi j Tim
In some sense, minimizing network-wide BFP can be viewed as minimizing a

weighted version of bandwidth usage, where the weight is the interference
109 5.4 Problem formulation
footprint area. By (5.4), (5.9) is equivalent to minimizing W (RImax )2

% m &2/
pij
iN mMi j T m
P . Since W (RImax )2 is a constant factor, we
i
can remove it from the objective function.
Discretization of transmission powers For power control, we assume that
the transmission power can only be tuned into a finite number of discrete lev-
els between 0 and P . This discretization helps the branch-and-bound process
run much faster due to reduced optimization space. This discretization is also
consistent to power control in the real world, where in many cases a radios
transmission power can only be tuned into a finite number of discrete levels.
To model this discrete version of power control, we introduce an integer param-
eter Q that represents the total number of power levels to which a transmitter
1
can be adjusted, i.e., 0, Q 2
P, Q P , . . . , P . Denote qijm {0, 1, 2, . . . , Q} as the
qijm
m , i.e., p m =
integer power level for pij ij Q P. Then (C-1 ), (C-2 ), and (5.8) can
be rewritten as follows:
% &
dij
qijm Qx m
ij , Qx m
ij (5.10)
R max
% &T
dkj
m
qkh Q 1 Qxijm (k Ijm , k = i, h Tkm ) (5.11)
RImax
s(l)=
j,d(l)=i % &
gij P m
fij (l) W log2 1 + q .
W Q ij
lL mMij
We can re-formulate (5.11) as follows. Note that by (C-3), there is at most one
m = 1. As a result, based on (5.11), there is at most one
h Tkm such that xkh
m m . Thus, (5.11) can be rewritten as
qkh > 0 in hT m qkh
k
% &
dkj
m
qkh Q 1 max Qxijm (k Ijm , k = i).
m R I
hTk
This reformulation will help reduce the number of constraints associated with
(5.11).
Mathematical formulation Putting together the objective function and all
the constraints for power control, scheduling, and flow routing, we have the
following problem formulation:
# $2
qijm
Minimize
Q
iN mMi j Tim

subject to xijm 1 (i N , m Mi ) (5.12)
j Ti m
% &
dij
qijm Qxijm 0 (i N , m Mi , j Tim ) (5.13)
RTmax
qijm Qxijm 0 (i N , m Mi , j Tim ) (5.14)
% &
dkj
m
qkh + 1 Qxijm Q
m RImax
hTk
(i N , m Mi , j Tim , k Ijm , k = i)
s(l)=
j,d(l)=i % & (5.15)
gij P m
fij (l) W log2 1+ q 0 (i N , j Ti )
W Q ij
lL mMij
(5.16)
fij (l) = r(l) (l L, i = s(l)) (5.17)
=s(l)
j j Tik=d(l)

fij (l) fki (l) = 0 (l L, i N , i = s(l), d(l)) (5.18)
j Ti kTi
xijm {0, 1}, qijm {0, 1, 2, . . . , Q} (i N , m Mi , j Tim )

fij (l) 0 (l L, i N , i = d(l), j Ti , j = s(l)),
where (5.13) and (5.14) come from (5.10). In this problem formulation,
W, gij , RTmax , RImax , P , , r(l), and Q are constants; qijm , xijm , and fij (l) are
optimization variables.
This optimization problem is in the form of a mixed-integer nonlinear pro-
gramming (MINLP), which is NP-hard in general [46]. In the next section, we
develop a solution procedure based on the branch-and-bound framework that
we described in Section 5.1.
5.5 A solution procedure
As discussed in Section 5.1, with the branch-and-bound framework, there are

several key components that are problem specific and need to be carefully
designed. These components are listed as follows and will be designed in Sec-
tions 5.5.1 and 5.5.3:
How to obtain a tight relaxation as well as a tight lower bound?

How to design a local search algorithm to find a feasible solution and an
upper bound?
How to select a partitioning variable?
5.5.1 Linear relaxation

During each iteration of the branch-and-bound process, we need a relaxation
technique to derive a lower bound for the optimal objective function value
111 5.5 A solution procedure
Figure 5.3 Illustration of

(qk1)2/ (q )2/
convex hull for a discrete term. k
um (q2)2/
ij
(q1)2/
(q0)2/
q0 q1 q2 qk1 qk
qm
ij
(see Steps 3 and 4 in Algorithm 5.1). Since an LP problem can be solved

in polynomial time, we aim to obtain a polynomially sized linear relaxation
for the underlying problem. For a nonlinear discrete term, we propose to
use a convex hull relaxation. This is done by introducing a new variable um ij
for the nonlinear discrete term (qijm )2/ . Suppose that qijm {q0 , q1 , . . . , qK },
where q0 (qijm )L < q1 < < qK (qijm )U . The convex hull of feasible
(qijm , um
ij )-coordinates (see Fig. 5.3) can be formulated as follows:
(qK )2/ (q0 )2/ m qK (q0 )2/ (qK )2/ q0

ij
um qij
qK q0 qK q0
(q )2/ (q ) 2/ q 2/ (q )2/ q
k k1 k k1 )
(q k k1
um q m
(1 k K).
ij
qk qk1 ij
qk qk1
Similarly, we can introduce a new variable vijm for the nonlinear discrete term

gij P m
log2 1 + W Q qij and construct corresponding convex hull constraints for
vijm . The details of this part are left in a homework exercise.
Denote x and q as the vectors of variables xijm and qijm , respectively. We thus
derive the following linear relaxation for Problem z:
1 m
Minimize u
Q2/ ij
iN mMi j Tim
subject to Convex hull constraints for um

ij (i N , m Mi , j Tim )

xij 1
m
(i N , m Mi )
j Tim
% &
dij
qijm Qxijm 0 (i N , m Mi , j Tim )
RTmax
qijm Qxijm 0 (i N , m Mi , j Tim )
% &
dkj
m
qkh + 1 max Qxijm Q (i N , m Mi , j Tim , k Ijm , k = i)
RI
hTkm
s(l)=
j,d(l)=i
fij (l) W vijm 0 (i N , j Ti )
lL mMij
Convex hull constraints for vijm (i N , m Mi , j Tim )

fij (l) = r(l) (l L, i = s(l))
j Ti
=s(l)
j k=
d(l)
fij (l) fki (l) = 0 (l L, i N , i = s(l), d(l))
j Ti kTi
ij , vij 0
um (i N , m Mi , j Tim )
m
fij (l) 0 (l L, i N , i = d(l), j Ti , j = s(l))

(x, q) z ,
where z is the set of all possible values of (x, q) in Problem z. For example,
1 for the original problem (Problem 1) is {(x, q) : 0 xijm 1, 0 qijm
Q}. The above linear relaxation formulation is an LP problem, which can be
easily solved and its objective value serves as a lower bound for Problem z.
5.5.2 Local search algorithm

Denote the relaxation solution as z . Note that z may not be feasible because
the x- and q-values in z may not be integers after relaxation. We now show
how to obtain a feasible solution z based on z . The objective of this feasible
solution z will serve as an upper bound for z .
To construct a feasible solution z , we use the same routing solution as that
in z , i.e., we let f = f. We then need to determine integer values for x and q
in z such that constraints (5.12)(5.15) hold and that the routing solution f is
feasible, i.e., (5.16) holds for each link i j .
Initially, we set each qijm = (qijm )L and each xijm = (xijm )L , where
[(qijm )L , (qijm )U ] and [(xijm )L , (xijm )U ] are the respective value-sets of qijm and
xijm in Problem z. Based on discussion in Section 5.5.3, the initial x- and
q-values satisfy constraints (5.12)(5.15). Next, we consider (5.16). Based on
m -values, we can compute the capacity gij P m
these qij mMi W log2 1 + W Q qij
s(l)=j,d(l)=i
of each link i j . The requirement on a link i j is lL fij (l).
If a links requirement is larger than its current capacity, we attempt to satisfy
(5.16) by increasing qijm under its value-set limitation (with necessary adjust-
ments on other variables value-sets to satisfy constraint (5.15)). In particu-
lar, we need to keep track of the maximum allowed transmission power. This
Algorithm 5.2 An algorithm to obtain x and q.

A local search algorithm
Initialization
1. Set qijm = (qijm )L and xijm = (xijm )L .

gij P m
2. Compute the capacity mMi W log2 1 + W Q ij and the flow
q
s(l)=j,d(l)=i
rate requirement lL fij (l) on each link i j .
Main iteration
3. If flow rate requirement on each link is satisfied, then a feasible
solution is found.
4. Otherwise, among all the links with flow rate requirements larger than
link capacities, identify a link i j that has the largest requirement.
//Try to increase the capacity for link i j as follows.
5. Increase qijm among the currently used bands in nonincreasing
order of qijm and under the limitation that qijm qijm .
6. If the achieved capacity on link i j is sufficient, then go to Step 3.
7. Otherwise, use an available but currently unused band in
nonincreasing order of qijm .

dij
8. For the selected band m, increase qijm under the constraint R max
T
Q qijm qijm to satisfy (5.13) and set xijm = 1 to satisfy (5.14).
9. * m = 0 for h T , h = j to satisfy (5.12), and let (q m )
Set xih +
i kh U
dkj
RImax Q for k Ijm , k = i, h Tkm to satisfy (5.15).
11. Otherwise, increase qijm among the currently used bands in
nonincreasing order of qijm and under the constraint qijm (qijm )U .
13. Otherwise, link i j cannot be satisfied and thus a feasible solution
cannot be found.
Note: A band m is available on link i j if for any transmitting node k using this
band, node j is not in its interference range.
is done in Step 9 of Algorithm 5.2, where if a new band m is used for link
i j , then it is necessary to pose a limit on each neighboring transmitter k
m ) -value for
such that the interference from k is negligible. As a result, the (qkh U
k Ijm , k = i, h Tkm may be decreased to satisfy (5.15). Suppose that each
updated (qkh m ) is no less than (q m ) . Then we have a new solution and we
U kh L
can calculate the objective value of this solution. If we can satisfy (5.16) for
all links, then we have a feasible solution. Otherwise (i.e., if (5.16) on any link
cannot be satisfied), we declare that we cannot find a feasible solution and thus
we set the objective value to . The details of this local search algorithm are
presented in Algorithm 5.2.
5.5.3 Selection of partitioning variables

If the relaxation error for a Problem z is not small, the gap between its lower
and upper bounds may be large, i.e., LB < (1 )U B. To narrow this gap,
we create two new subproblems z1 and z2 from Problem z, in the hope that
these two new problems will have smaller relaxation errors, and thus will yield
tighter bounds for the objective function. To generate Problems z1 and z2 , we
identify a partitioning variable based on its ascribed relaxation error (Step 11 in
Algorithm 5.1). The partitioning process terminates whenever we have LB
(1 )U B.
The partitioning variables include all the x- and q-variables. For our prob-
lem, we find that the x-variables are more important than the q-variables, in
terms of their impact on the objective value. Thus, instead of identifying a
partitioning variable based on its relaxation error only, it makes sense to con-
sider the x-variables before the q-variables. That is, we first select one of the
x-variables for partitioning, as applicable.
In particular, for the relaxation solution z , we choose an xijm -variable hav-
ing the largest relaxation error min{xijm , 1 xijm } among all the x-variables and
let its value-set in Problems z1 and z2 be {0} and {1}, respectively. Since the
resulting value-set for this xijm only has one element in each new problem,
this variable can be replaced by a constant. As a result, some constraints may
also be removed. It should be noted that the new value-set of xijm may narrow
other variables value-sets, based on constraints on (5.12)(5.15). That is, if
the new value-set of xijm is {0}, then we have qijm = 0 to satisfy (5.14). Thus,
we have (qijm )L = (qijm )U = 0. If the new value-set of xijm is {1}, then we have

m = 0 for h T , h = j to satisfy (5.12); q m dij
xih i ij max
RT
Q to satisfy (5.13);

dkj
and qkh R max Q for k Ij , k = i, h Tk to satisfy (5.15). Thus, we
m m
I
m ) = (x m ) = 0; (q m ) is updated by max (q m ) , dij
have (xih L ih U ij L ij L RTmax Q ;

m ) is updated by min (q m ) , dkj
and (qkh U kh U RImax
Q . If these updates make
the new value-set of a variable empty, then thecorresponding new subproblem
d
is clearly infeasible. One sample scenario is R max ij
Q > (qijm )U . Then the
T
updated (qijm )L > (qijm )U , which yields an empty value-set for qijm . As a result,
we can fathom an infeasible subproblem (i.e., remove it from the problem
list).
After we are done with examining all the x-variables, i.e., all the x-variables
are at a value 0 or 1 in the relaxation solution, then we select one of the
q-variables for partitioning. In particular, for the relaxation solution z , the
relaxation error of a discrete term qijm is
min{qijm qijm , qijm + 1 qijm };
ij = (qij )
the relaxation error of a nonlinear discrete term um m 2/ is
|um
ij (qij )
m 2/
|;

gij P m
and the relaxation error of a nonlinear discrete term vijm = log2 1 + W Q qij is
% &
m
v log 1 + gij P q m .
ij 2
W Q ij
Among these three types of relaxation errors, we identify the largest one and
choose the corresponding qijm as the partitioning variable. Assuming the value-
set of qijm in Problem z is {q0 , q1 , . . . , qK }, its value-set in Problems z1 and
z2 will be {q0 , q1 , . . . , qijm } and {qijm + 1, qijm + 2, . . . , qK }, respectively.
The new value-set of qijm may narrow other variables value-sets due to con-
straint (5.15). That is, if we increase the lower bound for one q-variable in the
summation in constraint (5.15), the upper bounds of other q-variables in this
summation may be decreased. If these updates make the new value-set of a
variable empty, then the corresponding new subproblem is clearly infeasible.
Again, we only keep feasible subproblems in the problem list.
A feasible subproblem has at least one solution that satisfies constraints
(5.12)(5.15). We can further verify that if we set each qijm = (qijm )L and each
xijm = (xijm )L , then constraints (5.12)(5.15) are satisfied. In particular, con-
straints (5.12)(5.14) hold due to updates in the partitioning process. Since
constraint (5.15) holds by a solution with some qijm values no less than (qijm )L
and some xijm values no less than (xijm )L , decreasing these values will not vio-
late (5.15).

In this section, we consider a randomly generated 20-node ad hoc network
with each node located in a 50 x 50 area. For ease of exposition, we nor-
malize all units for distance, bandwidth, rate, and power based on (5.1) and
(5.8) with appropriate dimensions. An instance of network topology is given
in Fig. 5.4 with each nodes location listed in Table 5.2. We assume that there
are |M| = 10 frequency bands in the network and each band has a bandwidth
of W = 50. At each node, only a subset of these frequency bands is available.
In the simulation, this is done by randomly selecting a subset of bands for each
node. Table 5.2 shows the available bands at each node. Within this network,
we assume that there are |L| = 5 user sessions, with source and destination
nodes chosen randomly. The rate of each session is randomly generated within
[10, 100]. Table 5.3 specifies an instance of each sessions source node, desti-
nation node, and rate requirement.
We assume that RTmax = 20, RImax = 40, and that the path-loss index
= 4. The threshold PT is assumed to be PT = W = 50. Thus, we
Figure 5.4 A 20-node ad hoc

50
network. 20 8
4 19
16
40 12
3 7
30
11 15
14
20 2
6
17
10
9
10
1 5
18
13
0
0 10 20 30 40 50
Table 5.2 Each nodes location and available frequency bands for the 20-
node network.
Node index Location Available bands
1 (10.5, 4.3) I, II, III, IV, V, VI, VII, VIII, IX, X

2 (1.7, 17.3) II, III, IV, V, VI, VII, X
3 (10.7, 30.8) I, III, IV, V, VI, VII, VIII, IX, X
4 (10.2, 45.3) I, III, IV, V, VI, VII, VIII, IX, X
5 (17.8, 4) I, II, V, VI, VII, VIII, IX
6 (17.2, 15.2) I, II, IV, VIII
7 (16.9, 30.8) I, II, III, IV, V, VI, VII, VIII, IX, X
8 (12.3, 47.3) I, III, IV, V, VII, VIII, IX
9 (28.2, 11.5) I, III, V, VII
10 (32.1, 13.8) I, II, III, IV, VI, VII, VIII, IX, X
11 (30.4, 25.6) I, II, III, V, VI, VIII, IX, X
12 (29.7, 36) I, II, III, IV, VI, VI
13 (41.7, 3.1) I, II, III, V, VI, VIII, IX, X
14 (47.5, 20) I, IV, V, VIII, IX, X
15 (43.3, 25.3) II, III, IV, V, VI, VII, VIII, IX, X
16 (44.1, 42.7) I, II, IV, VI, VII, VIII, IX, X
17 (49.6, 15.8) I, II, III, IV, V, VI, VII, VIII
18 (28.7, 2.5) I, II, III, VI, VII, VIII, IX, X
19 (28, 43.5) II, IV, V, VI, VIII
20 (5, 46.9) II, IV, V, VI, VII

RTmax
have PI = RImax PT W = 50
16 and the maximum transmission power P =
(RTmax ) PT W = 8 106 . We set
= 0.05, which guarantees that the obtained
solution is within 5% of optimality.
Table 5.3 Source node, destination node, and rate requirement of the five
sessions.
Session Source node Destination node Rate requirement
1 7 16 28
2 8 5 12
3 15 13 56
4 2 18 75
5 9 11 29
Figure 5.5 Objective value as a 12

function of Q.
10
8
Objective value
0
1 3 5 7 9 11 13 15
Q
5.6.2 Results
In this set of results, we apply the proposed solution procedure to the 20-node
network described above for different levels of power control granularity (Q).
Note that Q = 1 corresponds to the case where there is no power control, i.e.,
a node always uses its peak power P for transmission. When Q is sufficiently
large, power control approaches a continuum. Fig. 5.5 shows the results. First,
we note that the granularity of power control has a significant impact on the
optimal objective value. Comparing the case when there is no power control
(Q = 1) and the case of Q = 15, we find that there is nearly a 40% reduction
in the optimal objective value. Second, although the optimal objective value
is a nonincreasing function of Q, when Q becomes sufficiently large (e.g., 10
in this network setting), further increase in Q does not have much reduction
in the optimal objective value. This suggests that, for practical purposes, the
number of required power control levels does not need to be a large number.
The rest of our results in this section are for Q = 10. For transmission
power, we obtain the following results:
Figure 5.6 Flow routing for the

five sessions in the 20-node 50
20 8
network. 4 19 16
40
12
3 7
30
11 15
14
20
2 6 17
10
9
10
5
1
18 13
0
0 10 20 30 40 50
1
q9,11 = 3, 2
q12,16 = 4, 3
q7,12 = 3,
4 = 4,
q2,1 q2,1 = 4,
5 q5,18 = 2,
6
7
q17,13 = 4, 8 = 5, q 8
q8,3 14,17 = 1,
9
q1,5 = 1, q15,14
9
= 1,
q3,2 = 5.
10
The scheduling results are as follows:

1
x9,11 = 1, 2
x12,16 = 1, 3
x7,12 = 1,
x2,1 = 1,
4 x2,1 = 1,
5 x5,18 = 1,
6
7
x17,13 = 1, 8 = 1, x 8
x8,3 14,17 = 1, x1,5 = 1, x15,14 = 1,
9 9
10 = 1.
x3,2
The flow routing topology is shown in Fig. 5.6. The corresponding flow rates
are as follows:
f7,12 (1) = 28, f12,16 (1) = 28,
f8,3 (2) = 12, f3,2 (2) = 12, f2,1 (2) = 12, f1,5 (2) = 12,
f15,14 (3) = 56, f14,17 (3) = 56, f17,13 (3) = 56,
f2,1 (4) = 75, f1,5 (4) = 75, f5,18 (4) = 75,
f9,11 (5) = 29.
Note that a link may be used by multiple sessions. For example, link 2 1
is used by Sessions 2 and 4. As a result, the total data rate on link 2 1 is
f2,1 (2) + f2,1 (4) = 12 + 75 = 87.
The following observations on the numerical results show the close cou-
pling relationships between per-node power control and the upper layers. In
one observation (scheduling), we can see that links 8 3 and 14 17 are
active on the same band 8. This is feasible because the interference range at
8
node 14 is 22.49 under q14,17 = 1, which is smaller than 38.35 (the distance
between nodes 3 and 14). We note that if there is no power control, i.e., node
14 uses the peak transmission power, then node 3 will be in the interference
range of node 14, which is RImax = 40, and will lead to a scheduling conflict. In
119 5.8 Problems
another observation (routing), we see that, for Session 2 (node 8 to node 5), the
routing path is 8 3 2 1 5. Here, a shorter path 8 3 6 5 is
not used. This is because path 8 3 6 5 is interfered by transmissions
on other paths. The optimal solution tends to choose paths that are not close
to each other. Finally, for Session 3 (node 15 to node 13), the routing path is
15 14 17 13, while a shorter path 15 17 13 is not used. This
is because node 15 can use a smaller transmission power to transmit to a closer
neighboring node 14 with a smaller interference footprint. This allows link
1 5 to be active on the same band 9 with link 15 14.
5.7 Chapter summary
This is the first chapter of the second part of this book, which presents some
methods to develop (1 )-optimal solutions. In this chapter, we presented
the branch-and-bound framework and showed how it could be applied to solve
discrete and combinatorial optimization problems. Such problems are typi-
cally considered most difficult in nonconvex optimization, and the branch-and-
bound framework offers a general purpose and effective approach. The effec-
tiveness of branch-and-bound resides in the careful design of each component
in its framework, such as computation of a lower bound, local search of an
upper bound, and selection of partitioning variables (in the case of a minimiza-
tion problem). It should be noted that the worst-case complexity of a branch-
and-bound-based method remains exponential, although a judicious design of
each component could achieve reasonable computational times in practice.
In the case study, we considered a per-node power control problem for a
multi-hop CRN. This problem has a large design space that involves a tight
coupling relationship among power control, scheduling, and flow routing,
which is typical for a cross-layer optimization problem. We developed a math-
ematical model and a problem formulation, which is a mixed-integer nonlinear
programming (MINLP) problem. We showed how to apply the branch-and-
bound framework to design a solution procedure. Under the branch-and-bound
framework, we showed how to derive a linear relaxation to compute a lower
bound, how to perform a local search to find a feasible solution (upper bound),
and how to select suitable partitioning variables. Despite its worst-case expo-
nential complexity, the solution procedure that we developed here is a viable
approach to solve the per-node power control problem for a multi-hop CRN
(as demonstrated in the numerical examples).
5.8 Problems
5.1 Describe the branch-and-bound framework for an optimization problem

with a maximization objective (instead of minimization).
5.2 In the case study, we employed the so-called protocol model. Read
[152] and discuss the pros and cons of this model.
5.3 Show that for a session, if flow balance holds at the source node
[Eq. (5.5)] and at the relay nodes [Eq. (5.6)], then flow balance also holds
at this sessions destination node [Eq. (5.7)].
5.4 Referring to the RHS of (5.8), explain why there is no interference term
in the denominator inside the log function to calculate SINR.
5.5 In this chapter, we used the so-called network-wide bandwidth-footprint-
product (BFP) as the objective function. However, it is possible that two or
more neighboring nodes transmitting on the same frequency band may have
a partial overlap of their footprints. As a result, the overlapped area will be
counted multiple times in the objective function. Give a justification on why
this still makes sense.
5.6 In the formulation of the original problem, what constraints are associ-
ated with a single layer? What constraints couple multiple layers together?
5.7 In the formulation of the original problem, we discretize transmission
power from a continuous variable to a discrete variable. Then in the relax-
ation step of the branch-and-bound solution procedure, we change the discrete
transmission power variable to a continuous variable. Explain the purpose of
this back and forth change between continuous and discrete variables for the
transmission power.
5.8 We showed how to construct a linear relaxation for the discrete term
(qijm )2/ . Provide the details for deriving a linear relaxation for the discrete

gij P m
term log2 1 + W Q ij in the problem formulation.
q
5.9 There is another approach to obtain a linear relaxation for the discrete
term (qijm )2/ , where qijm {q0 , q1 , . . . , qK }. Introduce binary variables yijmk ,

where yijmk = 1 if qijm = qk and yijmk = 0 otherwise. Thus, K k=0 yij = 1 and
mk

qijm = K mk m
k=0 yij qk . We also introduce a new variable uij for (qij )
m 2/ and have
K
umij =
mk 2/ . A linear relaxation can be obtained by relaxing y mk
k=0 yij (qk ) ij
to a continuous variable in [0, 1]. Compare this approach and the linear relaxa-
tion approach in Section 5.5.1 and analyze which provides a tighter relaxation.
5.10 In local search, why do we initialize qijm at (qijm )L and increase it
upward? Comment on the strategy of starting qijm at (qijm )U and decreasing
it downward.
5.11 In local search, we focus on increasing the power values (see Steps 5,
8, and 11 in Algorithm. 5.2) to ensure that constraint (5.16) holds for each link
i j . Explain how other constraints in the original problem hold.
121 5.8 Problems
5.12 In the selection of partitioning variables, why are the x-variables more
important than the q-variables? After branching a problem into two new
subproblems, how can we revise the value-sets of the other x- and q-variables
by using constraints (5.12)(5.15)?
5.13 When we examine the relaxation errors for the q-variables, why can
ij (qij )
|um m 2/ | be greater than zero?
CHAPTER
6 Reformulation-linearization
technique and applications
Destiny is not a matter of chance, it is a matter of choice. It is not a thing to be

waited for, it is a thing to be achieved.
William Jennings Bryan
6.1 An introduction of reformulation-linearization

technique (RLT)
In Chapter 5, we presented a framework for branch-and-bound algorithms for
solving nonconvex programming problems. There are several key components
in this framework that are problem specific and need to be carefully custom-
designed. In particular, we need to develop a tight relaxation solution, i.e.,
one that yields a tight upper bound for a maximization problem (or a tight
lower bound for a minimization problem). For the case study in Chapter 5, we
presented two different linear relaxations for two particular types of nonlinear
terms (see Section 5.5.1). However, such relaxations may not be applicable to
some other nonlinear terms. Hence, we address this important question in this
chapter. That is, is there a relaxation approach that can be applied to a general
class of nonlinear terms?
In this chapter, we show that the reformulation-linearization technique
(RLT) [140; 141; 142; 146], developed by a co-author of this book (Prof. Hanif
Sherali) is such an approach. RLT is a systematic approach for deriving tight
linear relaxations for any monomial (i.e., a polynomial term of the form
n ci
i=1 (xi ) in variables xi , where the ci -exponents are constant integers; exten-
sions to rational exponents are presented in [144]).
The idea of RLT is best explained with a simple example. Suppose that
we have a nonlinear (bilinear) term x1 x2 in a nonconvex program involving
variables x1 and x2 . To derive a linear relaxation via RLT, we introduce a
122
123 6.1 An introduction of reformulation-linearizationtechnique (RLT)
new variable X{1,2} to represent x1 x2 and add suitable linear constraints for
this new variable. Suppose that we have derived (or are given) the lower
and upper bounds for variables x1 and x2 , i.e., (x1 )L x1 (x1 )U and
(x2 )L x2 (x2 )U , respectively. Then the following so-called bound-factor
constraints must hold:
[x1 (x1 )L ] [x2 (x2 )L ] 0,
[x1 (x1 )L ] [(x2 )U x2 ] 0,
[(x1 )U x1 ] [x2 (x2 )L ] 0,
[(x1 )U x1 ] [(x2 )U x2 ] 0.
Substituting X{1,2} for x1 x2 in the above constraints, we obtain
(x1 )L x2 + (x2 )L x1 X{1,2} (x1 )L (x2 )L , (6.1)
(x1 )U x2 + (x2 )L x1 X{1,2} (x1 )U (x2 )L , (6.2)
(x1 )L x2 + (x2 )U x1 X{1,2} (x1 )L (x2 )U , (6.3)
(x1 )U x2 + (x2 )U x1 X{1,2} (x1 )U (x2 )U . (6.4)
Therefore, a linear relaxation for the original problem can be obtained by
replacing the bilinear term x1 x2 by X{1,2} throughout the problem and adding
the above four linear constraints. Note that (6.1) (6.4) are valid restrictions
that relate the original variables x1 and x2 and the new RLT-variable X{1,2} .
In particular, X{1,2} = x1 x2 holds true when either x1 or x2 equals its corre-
sponding lower or upper bound value. This feature, which holds in a more
general form (see [141]), plays a decisive role in the tightness of RLT.
Next, we show how to obtain a linear relaxation for x12 in a nonconvex pro-
gram. Viewing x12 as x1 x1 , we derive a linear relaxation using RLT, by intro-
ducing a new variable X{1,1} to represent x12 , where the subscript of X repeats
the index 1 twice corresponding to the exponent in x12 , and by adding suitable
linear constraints for relating the new variable X{1,1} to the original variable
x1 . This is done similar to the above, by regarding x12 as a special case of x1 x2
when x2 x1 . Therefore, a linear relaxation for the original problem can be
obtained by replacing all occurrences of x12 within the polynomial program by
X{1,1} and by adding the following three bound-factor linear constraints. Note
that for this special case, both (6.2) and (6.3) become (6).6:
2(x1 )L x1 X{1,1} [(x1 )L ]2 , (6.5)
[(x1 )U + (x1 )L ] x1 X{1,1} (x1 )U (x1 )L , (6.6)
2(x1 )U x1 X{1,1} [(x1 )U ]2 . (6.7)
This strategy can be used in a likewise fashion for constructing a linear relax-

ation via RLT for a general monomial ni=1 (xi )ci in variables xi , where the
ci -exponents are integer constants for 1 i n. Although the general RLT
process is described in [141], we present a sequential quadrification process
[143] here for the sake of simplicity. This strategy adopts a two-step approach
for the linearization process. In the first step, we introduce new variables and
124 Reformulation-linearization technique and applications

identities to represent ni=1 (xi )ci using quadratic terms. In the second step, we
introduce RLT-variables as above to represent the quadratic terms and generate
the related RLT bound-factor constraints in order to derive a linear relaxation.
The following example illustrates the general idea:
Example 6.1
Suppose that we have a nonlinear term x1 x25 x3 in a nonconvex program
with variables x1 , x2 , and x3 . We begin by factoring x1 x25 x3 into quadratic
relationships as follows by defining new variables (note that this is not a
unique representation [143]):
X{2,2} = x22 , (6.8)
X{2,2,2,2} = 2
X{2,2} , (6.9)
X{1,2} = x1 x2 , (6.10)
X{1,2,2,2,2,2} = X{1,2} X{2,2,2,2} , (6.11)
X{1,2,2,2,2,2,3} = X{1,2,2,2,2,2} x3 . (6.12)
Note that in terms of the original variables x1 , x2 , and x3 , we have that
X{1,2,2,2,2,2,3} = x1 x25 x3 . Hence, we replace x1 x25 x3 in the given polyno-
mial program with the new variable X{1,2,2,2,2,2,3} . Furthermore, as before,
we include suitable linearized bound-factor constraints to represent the
quadratic relationships in the above equalities. For example, for (6.8), we
include the following three linear constraints, as in (6).5 (6).7 above:
2(x2 )L x2 X{2,2} [(x2 )L ]2 ,
[(x2 )U + (x2 )L ] x2 X{2,2} (x2 )U (x2 )L ,
(x2 )U x2 X{2,2} [(x2 )U ]2 .
The similar three linear constraints for representing (6.9) are left as a home-
work exercise (note that implied lower and upper bounds would need to be
derived for the variables X{1,2} , X{2,2,2,2} , and X{1,2,2,2,2,2} ). For (6.10), we
generate the following four bound-factor linear constraints as in (6.1)
(6.4):
(x1 )L x2 + (x2 )L x1 X{1,2} (x1 )L (x2 )L ,
(x1 )U x2 + (x2 )L x1 X{1,2} (x1 )U (x2 )L ,
(x1 )L x2 + (x2 )U x1 X{1,2} (x1 )L (x2 )U ,
(x1 )U x2 + (x2 )U x1 X{1,2} (x1 )U (x2 )U .
The similar four linear constraints for representing each of the relation-
ships (6.11) and (6.12) are relegated to the exercises. This produces a linear
relaxation for the original problem with five additional variables and 24
additional linear constraints.
125 6.2 Case study
For a more comprehensive discussion and surveys of RLT and related theo-
retical results, we refer the readers to [146; 148; 149]. In the rest of this chapter,
we will show how RLT can be applied to solve problems in a wireless network.
6.2 Case study: Capacity maximization for multi-hop cognitive

radio networks under the physical model
We study a capacity maximization problem under the physical model, which
is also called the signal-to-interference-and-noise ratio (SINR) model. In this
model, concurrent transmissions are allowed and interference (due to trans-
missions by nonintended transmitter) is treated as noise. SINR at a receiver
not only depends on the transmission power at the corresponding transmitter,
but also depends on the transmission power at other transmitters. A transmis-
sion is successful if and only if SINR at the receiver is greater than or equal
to a threshold. The achieved transmission capacity is a function of SINR (via
Shannon capacity formula).
Consider a multi-hop CRN where each node has access to a set of available
bands (likely heterogeneous). We are interested in how to maximize the rates
of a set of user communication sessions, with joint consideration at the phys-
ical layer (via power control), the link layer (via frequency band scheduling),
and the network layer (via flow routing). We give a mathematical characteri-
zation of these layers and formulate a mixed-integer nonlinear programming
(MINLP) problem. For this optimization problem, we first identify core opti-
mization variables and the core optimization space based on the physical sig-
nificance of the variables. We devise an algorithm using RLT in concert with
the branch-and-bound framework to obtain a (1 )-optimal solution.
As we learned in Chapter 5, although the branch-and-bound framework is
standard, many components within this framework need to be custom-designed
for the specific problem. For our problem, we develop the following compo-
nents. (1) By applying RLT, we develop a tight linear relaxation so as to obtain
a tight upper bound for our objective. (2) To compute the lower bounds, we
design a local search algorithm by analyzing and removing infeasibility in the
resulting linear relaxation solution. (3) For problem partitioning, we develop
a branching strategy based on the physical significance of the partitioning
variables. With these carefully designed components, the overall branch-and-
bound solution procedure is able to compute a (1 )-optimal solution much
faster than using a brute-force exhaustive search.
The remainder of this chapter is organized as follows. Section 6.3 gives a
mathematical characterization of power control, scheduling, and routing in
the SINR model for a multi-hop CRN. In Section 6.4, we perform prob-
lem reformulation and obtain a cleaner and more compact formulation. Sec-
tion 6.5 analyzes the core optimization space and describes the algorithm to
obtain a (1 )-optimal solution. Section 6.6 presents some numerical results.

6.3 Mathematical models
Denote N as the set of nodes of a multi-hop CRN. Each node i N senses

its environment and finds a set of available frequency bands Mi that it can
use, which may not be identical to those at other nodes. We assume that the
bandwidth of each frequency band (channel) is W . Denote M the union of

all frequency bands among all the nodes in the network, i.e., M = iN Mi .
(
Denote Mij = Mi Mj , which is the set of frequency bands that is common
between nodes i and j and thus can be used for communication between the
two nodes. In the rest of this section, we present mathematical modeling for
each layer and formulate a throughput maximization problem. Table 6.1 lists
the notation.
6.3.1 Power control, scheduling, and their relationship in the SINR model
Power control on each transmitting node at the physical layer affects SINR at
a receiving node. These SINR values in turn will affect scheduling decisions at
the link layer. That is, if a node is scheduled to receive, then its SINR must be
at least Smin (minimum threshold requirement). Therefore, power control and
scheduling are tightly coupled via SINR and cannot be modeled separately.
Scheduling at a node can be done either in the frequency domain or time
domain. In this chapter, we consider scheduling in the frequency domain in the
form of assigning frequency bands (channels). Note that a time domain-based
formulation can be done in a similar fashion.
In the SINR model, there may still be concurrent transmissions within
the same channel (and thus interference). Denote scheduling variables xijm as
follows:
'
1 if node i transmits data to node j on band m,
xijm =
0 otherwise.
We assume that a node can use a band for transmission (or reception) to (or
from) only one other node. That is,
j :mMj
k
k:mM
m
xki + xijm 1 (i N , m Mi ). (6.13)
kN ,k=i j N ,j =i
For power control, we assume that the transmission power at a node can
only be tuned to a finite number of levels between 0 and Pmax . To model this
discrete power control, we introduce an integer parameter Q that represents
the total number of power levels to which a transmitter can be adjusted,
127 6.3 Mathematical models
Table 6.1 Notation.
Symbol Definition
m
cij Capacity of link i j under pij m
d(l) Destination node of session l L

fij (l) Data rate for session l on link i j
gij Propagation loss from node i to node j
K Rate-scaling factor for all sessions
LBz Lower bound for Problem z
LB The maximum lower bound among all problems
L A set of user communication sessions in the network
Mi of available bands at node i N
A set
M = iN ( Mi , the set of frequency bands in the network
Mij = Mi Mj , the set of frequency bands on link i j
N A set of nodes in the network
m
pij The transmission power from node i to node j on band m
Pmax The maximum transmission power at a transmitter
m
qij Discrete transmission power level from node i to node j on band m
Q Number of discrete transmission power levels at a transmitter
r(l) Minimum rate requirement of session l
s(l) Source node of session l L
m
sij SINR from node i to node j on band m
Smin The minimum required SINR threshold
j :mMj m
tim = j N ,j =i pij , sum of transmission power at node i on band m
U Bz An upper bound for Problem z
UB The maximum upper bound among all problems
W Bandwidth of a frequency band
m
xij Binary variable to indicate whether or not band m is used on link i j
A small positive constant reflecting desired accuracy
z The core optimization space of the relaxed Problem z
z A local search solution for Problem z
A (1 )-optimal solution
i.e., the transmission power can be 0, Q 1 2

Pmax , Q Pmax , . . . , Pmax . Denote qijm
{0, 1, 2, . . . , Q} as the integer levels corresponding to their respective trans-
mission powers. Clearly, when node i does not transmit data to node j on band
m, qijm = 0. Thus, power control and scheduling are coupled with each other
via the following relationship:
)
[1, Q] if xijm = 1,
qijm (i, j N , i = j, m Mij ). (6.14)
=0 otherwise,
Consider a transmission from node i to node j on band m. When there is

interference from concurrent transmissions on the same band, the SINR at node
j , denoted as sijm , is
ij qm
gij Q Pmax
sijm = k:mMk h:mMh m
qkh
(i, j N , i = j, m Mij ), (6.15)
W + kN ,k=i hN ,h=k gkj Q Pmax
where is the ambient Gaussian noise density and gij is the propagation loss
from node i to node j .
Note that, in theory, for any small SINR, the corresponding capacity is still
positive (by Shannons capacity formula). But in practice, if SINR is too small,
then the achieved capacity will also be very small. In this case, such a weak
link will not be very useful to carry traffic flow. Thus, we may use a threshold
to remove such weak links from consideration. In this regard, we introduce
a threshold for SINR, i.e., a transmission from node i to node j on band m
is considered successful if and only if sijm Smin . We thus have the following
coupling relationship for scheduling (xijm ) and SINR (sijm ):
xijm = 1 sijm Smin (i, j N , i = j, m Mij ). (6.16)
6.3.2 Routing and link capacity

We assume that there is a set L of active user communication (unicast) ses-
sions in the network. Denote s(l) and d(l) as the source and destination nodes
of session l L, and r(l) as the minimum rate requirement (in b/s) for session
l. In our study, we aim to maximize a common scaling factor K for all ses-
sion rates. That is, we aim to determine the maximum K such that a rate of
K r(l) can be transported from s(l) to d(l) for each session l L. Again, for
optimality and flexibility, we allow flow splitting and multi-path routing inside
the network.
Mathematically, this can be modeled as follows. Denote fij (l) as the data
rate on link i j that is attributed to session l. If node i is the source node of
session l, i.e., i = s(l), then
j :Mij =

fij (l) = K r(l) (l L, i = s(l)). (6.17)
j N ,j =i
If node i is an intermediate relay node for flow attributed to session l, i.e.,

i = s(l) and i = d(l), then
j :Mij = ki =

k:M
fij (l) = fki (l) (l L, i N , i = s(l), d(l)).
j N ,j =i,s(l) kN ,k=i,d(l)
(6.18)

k:M ki =
fki (l) = K r(l) (l L, i = d(l)). (6.19)
kN ,k=i
129 6.4 Reformulation
In addition to the above flow balance equations at each node i N for ses-
sion l L, we impose a constraint to assure that the aggregated flow rates on
each radio link do not exceed this links capacity, i.e., for a link i j , we
have
s(l)=
j,d(l)=i
fij (l) W log2 (1 + sijm ) (i, j N , i = j, Mij = ).
lL mMij
(6.20)
This constraint shows the coupling relationship between flow routing and
SINR.
6.3.3 A throughput maximization problem

For throughput maximization, suppose we are interested in maximizing a
common scaling factor K for all sessions under some given minimum rate
requirements r(l). That is, we want to determine the maximum factor K
such that a rate of K r(l) can be transmitted from s(l) to d(l) for each
session l L in the network. Putting together the objective and all the con-
straints for power control, scheduling, and flow routing, we have the following
formulation:
Maximize K
subject to Constraints (6.13)(6.20)
xijm {0, 1}, qijm {0, 1, 2, . . . , Q}, tim , sijm 0 (i, j N , i = j, m Mij )
K, fij (l) 0 (l L, i, j N , i = j, i = d(l), j = s(l), Mij = ).
6.4 Reformulation
A formulation like the one in Section 6.3.3 is the first step in formulating our
cross-layer optimization problem. But it is in a rather raw form and more
work needs to be done to reformulate it into a more compact form that is
amenable to solution development. In this section, we analyze each constraint
in detail and perform some necessary and important reformulations.
The constraint described in (6.14) is not suitable for mathematical program-

ming. We reformulate it with the following linear constraint:
xijm qijm Qxijm (i, j N , i = j, m Mij ). (6.21)
It is easy to verify that this constraint is equivalent to (6.14).

Constraint (6).15 is in the form of a fraction. In a mathematical program, a

product form is more convenient to handle. We can rewrite (6).15 as follows:
qijm
Q Pmax
gij
sijm = k:mMk h:mMh qm
W + kN ,k=i hN ,h=k gkj Qkh Pmax
gij qijm
= W Q k:mMk h:mMh
(i, j N , i = j, m Mij ).
Pmax +
m
kN ,k=i hN ,h=k gkj qkh
This is equivalent to
W Q m k h:mM
k:mM h
s + m m
gkj qkh sij gij qijm = 0
Pmax ij
kN ,k=i hN ,h=k
(i, j N , i = j, m Mij ). (6.22)
m and s m are variables, while all other symbols are

Note that in (6.22), qkh ij
constants. Thus, we have a double sum of nonlinear terms qkhm s m in (6.22).
ij
To reduce the number of nonlinear terms, denote
h
h:mM
tkm = m
qkh (k N , m Mk ). (6.23)
hN ,h=k
Then (6.22) can be rewritten as
W Q m k
k:mM
sij + gkj tkm sijm gij qijm = 0 (i, j N , i = j, m Mij ), (6.24)
Pmax
kN ,k=i
which now only involves a single sum of nonlinear terms tkm sijm .
Similar to (6.14), the constraint described in (6.16) is not suitable for math-
ematical programming. It can be shown that (6.16) can be eliminated if we
have (6.21), (6).24, and the following new constraint:
sijm Smin xijm (i, j N , i = j, m Mij ). (6.25)
The verification of this fact is left as a homework problem.

Finally, we can easily prove that (6.17) and (6.18) imply (6.19). Thus, we can
remove constraint (6.19) and only retain (6.17) and (6.18) in the formulation.
With these careful reformulations, we now have a more compact problem

formulation, which is shown in Table 6.2.
Table 6.2 Problem formulation.
Maximize K
k:mMk m j :mMj m
subject to kN ,k=i xki + j N ,j =i xij 1 (i N , m Mi )
m xm 0
qij (i, j N , i = j, m Mij )
ij
m Qx m 0
ij
j :mMj m m
j N ,j =i qij ti = 0 (i N , m Mi )
W Q m k:mMk m m m
Pmax sij + kN ,k=i gkj tk sij gij qij = 0 (i, j N , i = j, m Mij )
m sm 0
Smin xij (i, j N , i = j, m Mij )
ij
j :Mij =
j N ,j =i fij (l) r(l)K = 0 (l L, i = s(l))
j :Mij = k:Mki =
j N ,j =i,s(l) fij (l) kN ,k=i,d(l) fki (l) = 0 (l L, i N , i = s(l), d(l))
s(l)=j,d(l)=i m
lL fij (l) mMij W log2 (1 + sij ) 0 (i, j N , i = j, Mij = )
m {0, 1}, q m {0, 1, 2, . . . , Q}, t m , s m 0

xij (i, j N , i = j, m Mij )
ij i ij
K, fij (l) 0 (l L, i, j N , i = j, i = d(l), j = s(l), Mij = )
6.5 A solution procedure
For the optimization problem in Table. 6.2, K, xijm , qijm , tim , sijm , and fij (l) are
optimization variables and Q, , W, Smin , Pmax , gij , and r(l) are constants.
This formulation is a mixed-integer nonlinear programming (MINLP), which
is NP-hard in general [46]. In Section 6.5.1, we first analyze the intricate rela-
tionship among the variables and identify the core variables among all the vari-
ables. We show that the dependent variables can be derived once these core
variables are fixed. We call the optimization space for the core variables the
core optimization space. In Section 6.5.2, we present the main algorithm on
how to determine an optimal solution in the core optimization space. Several
key components in the main algorithm are described in Sections 6.5.3, 6.5.4,
and 6.5.5.
6.5.1 Core variables

For the complex MINLP problem, its variables include xijm , qijm , tim , sijm , fij (l),
and K. However, a close investigation of these variables show that they are
inter-dependent. In particular, we find that the xijm - and qijm -variables are core
variables and the other variables {tim , sijm , fij (l), K} can all be derived based on
these core variables. How to derive these dependent variables based on the core
variables is left as a homework problem. As a result, we can focus our study
on an optimization space defined by the core variables, xijm and qijm , which is a
much smaller space.
6.5.2 A solution
In this section, we describe a solution procedure based on the branch-and-
bound framework. Recall that under branch-and-bound, we aim to provide a
(1 )-optimal solution, where is a small positive constant that reflects our
desired minimal accuracy (1 ) in computing the final solution. A branch-
and-bound framework for a minimization problem was presented in Sec-
tion 5.1. In this chapter, we need to solve a maximization problem, which
follows a similar scheme and is described in Algorithm 6.1.
Algorithm 6.1 A solution

Initialization
1. Let the initial best solution and the initial lower bound
LB .
2. Determine initial value-set for each core variable.
3. Initialize the problem list to include the original problem, and denote
this problem as Problem 1.
4. Obtain an upper bound U B1 for Problem 1.
Main iteration
5. Select Problem z that has the maximum U Bz -value among all
problems in the problem list.
6. Update upper bound U B = U Bz .
7. Find a feasible solution z along with a lower bound LBz .
8. If (LBz > LB) {
9. Update = z and LB = LBz .
10. If LB (1 )U B, stop with a (1 )-optimal solution .
11. Otherwise, remove all problems z with LB (1 )U Bz
from the problem list. }
12. Build two new problems z1 and z2 from Problem z.
13. Remove Problem z from the problem list.
14. Obtain U Bz1 and U Bz2 for Problems z1 and z2 .
15. If LB < (1 )U Bz1 , add Problem z1 to the problem list.
If LB < (1 )U Bz2 , add Problem z2 to the problem list.
16. Go to the next main iteration.
Since our core optimization space is finite (with a finite number of core
variables xijm and qijm , where each core variable has a finite integer value-set),
the branch-and-bound algorithm is guaranteed to converge in a finite number
of iterations. Several components (i.e., determining upper and lower bounds,
and partitioning subproblems) in the main algorithm are yet to be developed.
These components should exploit problem-specific structures to optimize per-
formance. In the rest of this section, we show how these components can be
designed.
6.5.3 Determining upper bounds

To find an upper bound for a subproblem in the proposed branch-and-bound
algorithm (see Steps 4 and 14 in Algorithm 6.1), we can construct a linear
relaxation by linearizing all the constraints in the model of Table 6.2. Hence,
the relaxed problem can be solved as an LP problem, where the optimal solu-
tion value provides an upper bound.
Note that in Table 6.2, tkm sijm and log2 (1 + sijm ) are nonlinear terms. For the
monomial tkm sijm , we can apply the RLT that we introduced in Section 6.1.
We define a new variable um m m
ij k to represent tk sij and introduce four additional
bound-factor linear constraints to relate this new variable to the original vari-
ables. In particular, suppose that tkm and sijm are bounded by (tkm )L tkm
(tkm )U and (sijm )L sijm (sijm )U , respectively. Then, this process yields the
following linear constraints for um ij k :
(tkm )L sijm + (sijm )L tkm um

ij k (tk )L (sij )L ,
m m
(tkm )U sijm + (sijm )L tkm um

ij k (tk )U (sij )L ,
m m
(tkm )L sijm + (sijm )U tkm um

ij k (tk )L (sij )U ,
m m
(tkm )U sijm + (sijm )U tkm um

ij k (tk )U (sij )U .
m m
For the nonlinear term log2 (1 + sijm ) = ln12 ln(1 + sijm ), we propose to
employ three tangential supports for ln(1 + sijm ), which yields a relaxation
for the underlying convex hull linear representation (see Fig. 6.1). Sup-
pose that sijm is bounded by (sijm )L sijm (sijm )U . We introduce a variable
m = ln(1 + s m ), and bound the convex hull of the region defined by the
cij ij
curve cijm over (s m ) s m (s m )
ij L ij ij U by using four segments as displayed
in Fig. 6.1, where the segments I, II, and III are tangential supports and
the segment IV is the chord. In particular, the three segments I, II, and III
are tangential at points (1 + (sijm )L , ln(1 + (sijm )L )), (1 + , ln(1 + )), and
(1 + (sijm )U , ln(1 + (sijm )U )), where
[1 + (sijm )L ] [1 + (sijm )U ] [ln(1 + (sijm )U ) ln(1 + (sijm )L )]

= 1
(sijm )U (sijm )L
Figure 6.1 A convex hull for

m = ln(1 + s m ). cm
ij
cij ij
III
II
cm=In(1+sm)
IV ij ij
sm +1
ij
(sm ) +1 b+1 (sm ) +1
ij L ij U
is the s-value corresponding to the intersection point of the segments I and III
(see Fig. 6.1); and segment IV is the chord that joins points (1 + (sijm )L , ln(1 +
(sijm )L )) and (1 + (sijm )U , ln(1 + (sijm )U )). The convex region defined by the
four segments can be described by the following four linear constraints:
[1 + (sijm )L ] cij
m s m [1 + (s m ) ] [ln(1 + (s m ) ) 1] + 1,
ij ij L ij L
m s m (1 + ) [ln(1 + ) 1] + 1,
(1 + ) cij ij
[1 + (sijm )U ] cij
m s m [1 + (s m ) ] [ln(1 + (s m ) ) 1] + 1,
ij ij U ij U
[(sijm )U (sijm )L ] cij

m + [ln(1 + (s m ) ) ln(1 + (s m ) )] s m
ij L ij U ij
(sijm )U ln(1 + (sijm )L ) (sijm )L ln(1 + (sijm )U ).
As a result, the nonlinear logarithmic term is relaxed using linear constraints.

After relaxing all such nonlinear terms for a subproblem in the branch-and-
bound process, say Problem z, we obtain a relaxed Problem z as formulated in
Table 6.3, which is an LP problem. In Problem z, x and q are vectors that
represent all the xijm - and qijm -variables, respectively; (xijm )L , (xijm )U , (qijm )L ,
and (qijm )U are constant bounds; z {(x, q) : (x)L x (x)U , (q)L q
(q)U } is the core optimization space of (x, q); and the bounds on the sijm -
variables for constructing the linearization are derived from the constraints of
the problem contingent on (x, q) z .
Any such relaxed problem z can be solved in polynomial time (since it is an
LP problem), where the optimal objective value provides an upper bound for
Problem z. We denote by LP (z) the corresponding optimal solution obtained
for the relaxed problem z.
6.5.4 Determining lower bounds

To compute a lower bound for Problem z, it is sufficient to find a feasible solu-
tion to this problem. By a feasible solution, we mean one that simply satisfies
all the constraints for Problem z in Table 6.2. Although any feasible solution
Table 6.3 Linear relaxation.
Maximize K
k:mMk m j :mMj m
subject to kN ,k=i xki + j N ,j =i xij 1 (i N , m Mi )
m xm 0
ij
m Qx m 0
ij
j :mMj m m
j N ,j =i qij ti = 0 (i N , m Mi )
W Q m k:mMk m m
Pmax sij + kN ,k=i gkj uij k gij qij = 0 (i, j N , i = j, m Mij )
Linear constraints for um

ij k (i, j, k N , i = j, m Mij , m Mk )
m sm 0
Smin xij (i, j N , i = j, m Mij )
ij
j :Mij =
j N ,j =i fij (l) r(l)K = 0 (l L, i = s(l))
j :Mij = k:Mki =
j N ,j =i,s(l) fij (l) kN ,k=i,d(l) fki (l) = 0 (l L, i N , i = s(l), d(l))
s(l)=j,d(l)=i W m
lL fij (l) mMij ln 2 cij 0 (i, j N , i = j, Mij = )
m
Linear constraints for cij (i, j N , i = j, m Mij )
tim , sij
m , cm , um 0
ij ij k (i, j, k N , i = j, m Mij , m Mk )
K, fij (l) 0 (l L, i, j N , i = j, i = d(l), j = s(l), Mij = )
(x, q) z
to Problem z can serve to provide a lower bound, it is desirable to find one that
offers a tight lower bound, i.e., with an objective value close to the optimal
value. Such a feasible solution (denoted by z ) can be found by searching the
neighborhood of LP (z), a process that is called local search.
A local search algorithm begins with an initial feasible solution. Such a solu-
tion may be far away from the optimum and may not provide a tight lower
bound. But we can iteratively improve the current solution to achieve a bet-
ter lower bound until we can no longer improve (increase) the lower bound,
whence we terminate this search process.
To obtain an initial feasible solution, we set xijm = (xijm )L for scheduling
and qijm = (qijm )L for power control. Then we can compute SINR value sijm
by (6).15. When an SINR value is larger than or equal to Smin , the achieved
capacity is W log2 (1 + sijm ). Otherwise (i.e., SINR < Smin ), the transmission is
Algorithm 6.2 Pseudocode for a local search algorithm

Initialization
1. Set xijm = (xijm )L and qijm = (qijm )L .
2. Compute the ratio ij based on (6.26) for each link i j , and denote
min = min{ij : i, j N , i = j, Mij = }.
Main iteration
3. Select a link i j such that ij = min .
4. If we can increase qijm on a used band {
5. Suppose that band m has the largest qijm -value in the solution
LP (z) among these bands.
6. Increase qijm such that qijm (qijm )U and the newly updated
kh > min for all other links k h. }
7. else, if we can increase qijm on an available and unused band {
8. Suppose that band m has the largest qijm -value in the solution
LP (z) among these bands.
9. Increase qijm such that qijm (qijm )U and the newly updated
kh > min for all other links k h.
10. If qijm increases, then set xijm = 1. }
11. else the iteration terminates.
considered unsuccessful. Note that although the flow rates fij (l) in the relaxed
solution LP (z) guarantee flow balance at each node, such flow rates may
exceed the capacities on some links under the initial xijm - and qijm -values. To find
feasible flow rates for the current xijm - and qijm -values, we compare the achiev-
able link capacity (under this current solution) to the aggregated flow rates
fij (l) on each link i j by computing the ratio between the two (denoted as
ij ) as follows:

mMij W log2 (1 + sij )
m
ij = s(l)=j,d(l)=i . (6.26)
lL fij (l)
If ij < 1 for some link i j , then the aggregated flow rates exceed the
link capacity and the link capacity constraint on i j is violated. In this
case, we need to scale down the flow rates on link i j (to satisfy the link
capacity constraint) and the flow rates on all other links (to maintain flow
balance in the network) by a value ij . On the other hand, we want to
have a as large as possible so as to maximize the scaling factor (our objec-
tive). Such a value is the bottleneck value ij among all links (denoted as
min = min{ij : i, j N , i = j, Mij = }). We now have a complete solu-
tion min fij (l), (xijm )L , (qijm )L for routing, scheduling, and power control,
respectively. The achieved objective is min K, where K is the objective value
in the relaxed solution LP (z).
In the next iteration, we aim to improve the current solution. Note that if
we can increase min , then the current solution is improved. Suppose that link
i j is the link with ij = min . To increase ij , we try to increase the trans-
mission power qijm on some band m under the constraint qijm (qijm )U . Based
on the constraints in Table 6.2, we may update the values of other variables to
maintain feasibility. For example, by the first constraint in Table 6.2, we need
to increase xijm from 0 to 1 if qijm is increased to a positive value. As a con-
sequence of increasing qijm , the interference with other transmissions on band
m is increased and thus the achieved capacities on other links are decreased.
Thus, qijm can be successfully increased only if for any other link k h, its
updated kh does not fall below the current min . If the current solution can
be improved (with a larger min ), then we continue to the next iteration and
try for further improvement. Otherwise, the local search algorithm terminates.
The pseudocode for this local search algorithm is given in Algorithm 6.2.
6.5.5 Partitioning approach

In a standard branch-and-bound procedure, partitioning (see Step 12 in
Algorithm 6.1) is done by choosing a variable having the largest relaxation
error and its value in the relaxed solution LP (z) is used to split its value-set
into two smaller sets. The reason for selecting a variable having the largest
relaxation error is that such a variable is most likely to influence the gap
between the upper and lower bounds. Thus, by partitioning its value-set, the
relaxation error might become smaller. This branching process also partitions
the optimization space for Problem z into two subspaces, thus resulting in two
new problems z1 and z2 , respectively.
Such a partitioning technique is of general purpose and does not exploit any
problem-specific property that could be used to prioritize the partitioning vari-
ables. We find that if we weigh the significance of each variable when choosing
a partitioning variable, the complexity of the overall algorithm can be signif-
icantly decreased. In particular, for our problem, we find that the x-variables
are more important than the q-variables. So we should first decide whether
or not a band is used for transmission. Only if a band is used can we further
explore the transmission power on this band. Thus, we first choose some suit-
able x-variable for partitioning, and when none of them remains (i.e., all the
x-variables are either 0 or 1 in the relaxation solution), we then consider the
q-variables for partitioning. Incidentally, we note that commercial software
(e.g., CPLEX) also provide the opportunity for users to specify priorities
among potential partitioning variables.
When choosing a specific xijm -variable for partitioning, we select one having
the maximum relaxation error, which is defined as min{xijm , 1 xijm }, where xijm
is the relaxed solution value for xijm in LP (z). Note that this error is positive
if and only if the variable is not presently binary-valued. Once partitioned,
the value-set for this variable is set to either 0 or 1 in the two respective
subproblems.
We further observe that by fixing the value of xijm , some other variables may
also be fixed. Note that by (6.13), a node can only receive from or transmit to
one node on the same band. Based on this observation, if the value of xijm is
set to 1, then we have xki m = 0 for k N , k = i, if m M ; x m = 0 for p
k ip
N , p = j , if m Mp ; xjmh = 0 for h N , h = j , if m Mh ; and xqj m =0
for q N , q = i, if m Mq . On the other hand, if the value of xij is set to 0,

m
then we have qijm = 0 based on (6.14) (or (6.21)).

When we are done with all the x-variables, we then consider partitioning
on the q-variables. Similar to what we did on the x-variables, we select a
q-variable that has the maximum relaxation error for partitioning. The relax-
ation error of a variable qijm is defined as min{qijm qijm , qijm + 1 qijm },
where qijm is the value of qijm in the relaxed solution LP (z). Again, note that
this relaxation error is positive if and only if qijm is not binary-valued. The
value-set of qijm in Problem z is {(qijm )L , (qijm )L + 1, . . . , (qijm )U }. Hence, its
new value-set in the two subproblems will be {(qijm )L , (qijm )L + 1, . . . , qijm }
and {qijm + 1, qijm + 2, . . . , (qijm )U }, respectively.
In this section, we present some numerical results for the solution procedure.

We consider 20-, 30-, and 50-node CRNs with each node randomly located
in a 50x50 area (see Tables 6.4, 6.6, and 6.11). For ease of exposition, we
normalize all units for distance, bandwidth, rate, and power based on (6.20)
with appropriate dimensions. We assume that for the 20-, 30-, and 50-node
CRNs, there are |M| = 10, 20, and 30 frequency bands in the network and
that each band has a bandwidth of W = 50. At each node, only a subset of
these bands is available. For the 20- and 30-node CRNs, we assume that there
are five user communication sessions (see Tables 6.5 and 6.7) and for the
50-node CRN, we assume that the number of user communication sessions
is 10 (see Table 6.12). The source node and destination node for each session
are randomly selected, and the minimum rate requirement of each session is
randomly generated within [1, 10].
We assume that gij = dij4 and that the SINR threshold Smin = 3 [54]. We
assume that under the maximum transmission power, a node at a distance of
4
20 can receive data when there is no interference. Thus, we have (20) WPmax =
Smin , i.e., the maximum transmission power is Pmax = Smin (20)4 W =
4.8 105 W . We assume that power control can be done at Q = 10 levels.
Table 6.4 Location and available frequency bands at each node for
a 20-node network.
Node Location Available bands
1 (0.1, 9.9) 1, 2, 3, 4, 7, 8, 9, 10
2 (29.2, 31.7) 1, 2, 3, 4, 5, 7, 8, 10
3 (3, 31.1) 1, 4, 5, 6
4 (11.8, 40.1) 1, 2, 3, 4, 6, 9, 10
5 (15.8, 9.7) 1, 2, 3, 5, 6, 8, 9
6 (16.3, 19.5) 3, 5, 6, 8, 9
7 (0.6, 27.4) 1, 4, 8, 9, 10
8 (22.6, 40.9) 1, 2, 3, 5, 7, 9, 10
9 (35.3, 10.3) 2, 9
10 (31.9, 19.6) 1, 2, 3, 4, 5, 6, 7, 8, 9, 10
11 (28.1, 25.6) 1, 2, 3, 4, 5, 6, 7, 8, 9, 10
12 (32.3, 38) 1, 8, 9, 10
13 (47.2, 2.6) 3, 5, 10
14 (44.7, 15) 2, 3, 6, 7, 8
15 (44.7, 24) 1, 2, 3, 4, 5, 6, 7, 8, 9, 10
16 (47.9, 43.8) 1, 3
17 (46.4, 16.8) 1, 7, 9
18 (11.5, 12.2) 2, 5, 6, 10
19 (28.2, 14.8) 4, 5, 6, 7, 8, 9, 10
20 (2.5, 14.5) 1, 7, 10
For our proposed algorithm, we set 0.1, which guarantees that the solu-
tion is at least 90% optimal.
We note that the brute-force approach cannot solve the problem even for
the 20-node CRN. The solution space of the capacity problem in Section 6.3.3
includes all possible sets of values for (xijm , qijm , K, fij (l)). Thus, the number
of solutions examined in the brute-force approach is clearly more than the
number of all possible sets of values for the qijm -variables. For the 20-node
CRN, the number of qijm -variables is about 20 (20 1) 5 = 1900, where 5
is an approximation of the average number of available bands on a link. Each
qijm -variable has (Q + 1) or 11 possible values. Thus, the number of all possi-
ble sets of values for the qijm -variables is about 111900 . Therefore, for the 20-
node CRN, the number of solutions examined in the brute-force approach is
at least 111900 . Even if each solution can be examined in 106 second, the run
time is 111900 106 > 111894 seconds, which is 111894 /(365 24 60 60) >
111894 /108 > 111886 years! Therefore, a brute-force approach cannot be used
to solve our problem except for toy-sized instances.
6.6.2 Results
For the 20-node network with five sessions, our solution achieves a scal-
ing factor K = 13.24. Based on the minimum rate requirement r(l) in
Table 6.5 Source node, destination node, and minimum rate requirement
for each session in the 20-node network.
Session l Source node s(l) Dest. node d(l) Min. rate req. r(l)
1 16 10 9
2 18 3 1
3 12 11 4
4 13 17 3
5 15 6 2
Figure 6.2 The routing 50

topology for the 20-node 16
45
five-session network. 4 8
40 12
35
3 2
30
7
25 11
15
6
20 10
17
20
15 18 14
19
10 1 5 9
5
13
0
0 5 10 15 20 25 30 35 40 45 50
Table 6.5, the corresponding flow rates K r(l) for the five sessions are
119.16, 13.24, 52.96, 39.72, 26.48, respectively. Fig. 6.2 shows the routing
topology for the final solution. The flow rates for each session on the links
along its path are as follows:
Session l = 1: f16,12 (1) = 119.16, f12,8 (1) = 103.30, f12,11 (1) = 15.86,
f8,2 (1) = 103.30, f11,10 (1) = 15.86, f2,10 (1) = 103.30;
Session l = 2: f18,20 (2) = 13.24, f20,1 (2) = 13.24, f1,7 (2) = 13.24,
f7,3 (2) = 13.24;
Session l = 3: f12,11 (3) = 52.96;
Session l = 4: f13,14 (4) = 39.72, f14,17 (4) = 39.72;
Session l = 5: f15,19 (5) = 26.48, f19,6 (5) = 26.48.
It is easy to verify that flow balance holds at all nodes. Note that flow splitting
and multi-path routing are used for Session 1, which has the largest rate
requirement.
Our solution also prescribes the scheduling variables xijm as follows, where
we indicate only the nonzero xijm -variables:
Band m = 1: x7,3 1 = 1, x 1
16,12 = 1;
Band m = 2: x8,2 2 = 1;
Band m = 3: x13,14
3 = 1;Band m = 4: x1,7
4 = 1, x 4
2,10 = 1;
Band m = 5: x11,10
5 = 1;
Band m = 6: x15,19 = 1;
6
Band m = 7: x14,17
7 = 1, x20,1
7 = 1;
Band m = 8: x12,11 = 1;
8
Band m = 9: x12,8
9
= 1, x19,6
9
= 1;
Band m = 10: x18,20 = 1.
10
The transmission power levels on the respective frequency bands are as

follows:
1 = 1, q 1
Band m = 1: q7,3 16,12 = 7;
Band m = 2: q8,2
2 = 2;
Band m = 3: q13,14
3 = 2;
Band m = 4: q1,7
4 = 7, q 4
2,10 = 2;
Band m = 5: q11,10 = 1;
5
Band m = 6: q15,19
6 = 9;
Band m = 7: q14,17
7 = 1, q20,1
7 = 1;
Band m = 8: q12,11 = 3;
8
Band m = 9: q12,8
9
= 1, q19,6
9
= 3;
Band m = 10: q18,20 = 1.
10
Note that the same frequency band may be used by concurrent transmissions.
1 = 1 and x 1
For example, since x7,3 16,12 = 1, we have that both nodes 7 and
16 are transmitting on band 1. Such concurrent transmissions are allowed as
long as the SINR at each receiving node is no less than Smin . For example,
g q 1
the SINR at the receiving node 12 on band 1 is s16,121 = W Q16,12 16,12 1 =
Pmax +g7,12 q7,3
(1.303105 )7
= 4.22, which is larger than Smin = 3. Thus, the trans-
2.083105 +(8.011107 )1
mission 16 12 on band 1 is successful. Following the same token, we can
compute all SINR sijm -values as follows:
Band m = 1: s7,31 = 118.47, s 1
16,12 = 4.22;
Band m = 2: s8,2 = 5.84;
2
Band m = 3: s13,14
3 = 3.75;
Band m = 4: s1,74 = 3.33, s 4
2,10 = 3.14;
Band m = 5: s11,10 = 18.87;
5
Band m = 6: s15,19
6 = 3.39;
Band m = 7: s14,17 = 1261.14, s20,1
7 7 = 65.46;
Band m = 8: s12,11 = 4.90;
8
Band m = 9: s12,8
9
= 3.56, q19,6
9
= 4.74;
Band m = 10: s18,20 = 6.45.
10
As expected, we see that the achieved SINR at each receiving node on each
band is larger than Smin = 3 in the derived solution.
For each link, we can further verify that the flow rates on this link do
not exceed its capacity. For example, for link 16 12, there is a flow rate
f16,12 (1) = 119.16 on this link. The achieved capacity is W log2 (1 + s16,12
1 )=
W log2 (1 + 4.22) = 119.21.
The above results are for the 20-node network. The results for the 30-
and 50-node networks are similar and we abbreviate our discussion. For
the 30-node five-session network, the available bands at each node and the
location of each node are displayed in Table 6.6. The source, destination,
and minimum rate requirement for each session are shown in Table 6.7.
The routing topology for the derived solution is depicted in Fig. 6.3. The
achieved scaling factor is 31.18. Accordingly, based on the minimum rate
requirement r(l) in Table 6.7, the flow rates K r(l) for the five sessions
Table 6.6 Location and available frequency bands at each node for a
30-node network.
1 (7, 0.7) 1, 2, 6, 7, 16, 17, 19, 20

2 (5, 4) 3, 5, 9, 12, 14, 15
3 (6.8, 14) 1, 2, 6, 7, 8, 11, 16, 17, 19, 20
4 (15.7, 3.3) 1, 2, 7, 16, 20
5 (9.5, 17) 3, 4, 5, 9, 12
6 (19.4, 17.1) 1, 2, 6, 7, 8, 16, 19, 20
7 (34.7, 14.6) 3, 4, 5, 9, 12, 14
8 (4.9, 25.9) 3, 4, 12
9 (46.6, 42.1) 10, 18
10 (8.3, 38.3) 3, 4, 5, 9, 14
11 (26.7, 11.1) 1, 6, 7, 8, 11, 16, 17, 19, 20
12 (36.4, 47.3) 10, 13, 18
13 (24.3, 21.2) 1, 2, 6, 8, 11, 19
14 (23.1, 0.8) 3, 5, 9, 14
15 (21.4, 19.2) 4, 9, 12, 14
16 (30.3, 28.1) 7, 8, 11, 16, 17, 19, 20
17 (32, 41.1) 7, 11, 16, 17, 19, 20
18 (14.1, 33.7) 3, 4, 5
19 (23, 46.4) 3, 12, 15
20 (30.3, 9.3) 5, 9
21 (17.6, 29.2) 1, 2, 6, 7, 8, 11, 16, 17, 19, 20
22 (27.1, 27.8) 9, 12, 14, 15
23 (26.9, 45.9) 3, 4, 5, 9, 10, 12, 13, 14, 15, 17
24 (43.3, 32.4) 1, 2, 11, 16, 17, 20
25 (45.4, 8.2) 3, 4, 5, 9, 12, 14
26 (43.4, 35) 3, 5, 9, 15
27 (41.3, 45.1) 1, 16, 20
28 (14.4, 30.3) 1, 2, 6, 7, 8, 11, 16, 17, 20
29 (41.6, 41.7) 3, 4, 5, 9, 10, 12, 14, 15, 18
30 (25.9, 12) 1, 2, 6, 7, 8, 11, 16, 17, 19, 20
for each session in the 30-node network.
1 16 28 4
2 24 11 7
3 13 1 1
4 19 29 8
5 26 15 1
Table 6.8 Flow rates for each session in the 30-node five-session network.
Session Session rate Flow rates of session l on each link

l K r(l) fij (l)
1 124.72 f16,21 (1) = 124.72, f21,28 (1) = 124.72;

f24,16 (2) = 113.91, f24,17 (2) = 104.36, f17,16 (2) = 104.36,
2 218.26 f16,21 (2) = 24.22, f16,13 (2) = 194.05, f21,28 (2) = 24.22,
f28,13 (2) = 24.22, f13,11 (2) = 160.59, f13,30 (2) = 57.68,
f30,11 (2) = 57.68;
3 31.18 f13,30 (3) = 31.18, f30,4 (3) = 31.18, f4,1 (3) = 31.18;
4 249.44 f19,23 (4) = 167.49, f19,29 (4) = 81.95, f23,26 (4) = 81.26,
f23,29 (4) = 86.23, f26,29 (4) = 81.26;
5 31.18 f26,22 (5) = 31.18, f22,15 (5) = 31.18.

23 12
19 27
five-session network. 9
40 17 29
10
35 26
18
30 21 24
28 22
8 16
25
13
20 15
5 6
15 7
3
30
10 20
11
25
5
2
4 14
0 1
0 5 10 15 20 25 30 35 40 45 50
are 124.72, 218.26, 31.18, 249.44, 31.18, respectively. The flow rates for each
session on links along its path are shown in Table 6.8. The nonzero schedul-
ing variables xijm and the nonzero power control variables qijm are specified in
Tables 6.9 and 6.10, respectively.
For the 50-node ten-session network, the available bands at each node and
the location of each node are displayed in Table 6.11. The source, destination,
and minimum rate requirement for each session are specified in Table 6.12.
Table 6.9 Scheduling for the 30-node five-session network.
Band Scheduling Band Scheduling Band Scheduling

1 = 1, x 1 8 15
1 x4,1 21,28 = 1; 8 x16,13 = 1; 15 x23,26 = 1;
2 2
x28,13 = 1; 9 9
x26,22 = 1; 16 16 16 = 1;
x24,17 = 1, x30,11
3 3
x19,29 = 1; 10 10 = 0;
xij 17 17
x24,16 = 1;
4 4
x23,29 = 1; 11 11
x17,16 = 1; 18 18 = 0;
xij
5 5
x26,29 = 1; 12 12
x19,23 = 1; 19 19
x13,30 = 1;
6 6
x13,11 = 1; 13 13 = 0;
xij 20 20 = 1.
x30,4
7 7
x16,21 = 1; 14 14
x22,15 = 1;
Table 6.10 Transmission power levels for the 30-node five-session network.
Band Transmission power Band Transmission power

1 = 1, q 1 11
1 q4,1 21,28 = 1; 11 q17,16 = 2;
2 2
q28,13 = 3; 12 12
q19,23 = 1;
3 3
q19,29 = 9; 13 13 = 0;
qij
4 4
q23,29 = 4; 14 14
q22,15 = 1;
5 5
q26,29 = 1; 15 15
q23,26 = 10;
6 6
q13,11 = 2; 16 16
q24,17 16
= 3, q30,11 = 1;
7 7
q16,21 = 4; 17 17
q24,16 = 7;
8 8
q16,13 = 2; 18 18 = 0;
qij
9 9
q26,22 = 7; 19 19
q13,30 = 1;
10 10 = 0;
qij 20 20 = 4.
q30,4
The routing topology for the derived solution is depicted in Fig. 6.4. The
achieved scaling factor is 13.36. The detailed solution regarding power control,
scheduling, and routing are similar to the 20- and 30-node networks, and are
hence omitted.
6.7 Chapter summary
When applying a branch-and-bound framework to solve nonconvex program-

ming problems, it is necessary to develop a tight relaxation, i.e., one that yields
a tight upper bound for a maximization problem (or a tight lower bound for a
minimization problem). In the last chapter (see Section 5.5.1), we showed two
different linear relaxations for two particular types of nonlinear terms. How-
ever, such relaxations may not be applicable to some other nonlinear terms.
In this chapter, we described the reformulation-linearization technique
(RLT) to derive tight linear relaxations for any monomial. Simply put, RLT
Table 6.11 Location and available frequency bands at each node for a 50-node
network.
1 (11.1, 21.7) 2, 3, 4, 8, 25
2 (0.1, 4) 6, 7, 10, 13, 14, 20, 23, 24, 26, 28
3 (7.2, 16.6) 6, 10, 14, 20, 23, 24, 26
4 (11, 32.2) 6, 7, 10, 13, 14, 20, 23, 24, 26, 28
5 (16.3, 3.6) 10, 13, 14, 20, 23
6 (14.5, 24.7) 8, 11, 25
7 (14.9, 13.7) 5, 9, 12, 16, 17, 18, 22, 27, 29, 30
8 (19.5, 14.9) 7, 24, 28
9 (26.6, 13.4) 1, 19, 21, 25
10 (22.5, 29.3) 1, 3, 4, 8, 11, 15, 19
11 (24.6, 40.5) 3, 8, 25
12 (38.4, 13.1) 2, 8, 11, 15
13 (4, 3.9) 9, 12, 16, 22, 27, 29, 30
14 (6.1, 18.6) 9, 12, 16, 17, 18, 22, 27, 30
15 (38.5, 22.6) 2, 4, 11, 15, 19, 21, 25
16 (1.2, 24.3) 5, 9, 12, 17, 22, 29, 30
17 (4.9, 42.3) 5, 27
18 (18.5, 1.4) 5, 9, 12, 17, 18, 27, 30
19 (16.9, 29.1) 3, 4, 10, 11, 12, 15
20 (33.5, 10.4) 7, 13, 14, 20, 23, 24, 26, 28
21 (25.6, 12.8) 6, 7, 20, 23, 24, 28
22 (45.2, 45.5) 2, 8, 15, 19
23 (43.6, 22.7) 1, 2, 3, 4, 11, 15, 19, 21
24 (10.6, 40.5) 4, 15, 19, 21, 25
25 (18.2, 32.7) 9, 12, 18, 22, 27
26 (25.2, 27.2) 10, 14, 20, 24, 26
27 (22.5, 42.2) 5, 9, 12, 16, 18, 27, 29, 30
28 (30, 31.5) 6, 13, 24, 26, 28
29 (35, 22.1) 6, 10
30 (25.7, 6.2) 5, 9, 12, 17, 18, 22, 27, 29, 30
31 (34.1, 12.4) 9, 12, 16, 17, 30
32 (26.4, 30) 5, 9, 12, 16, 17, 18, 22, 27, 29, 30
33 (14.1, 40.7) 1, 2, 25
34 (34.4, 46.5) 9, 17, 18, 30
35 (19, 22.5) 1, 6, 7, 10, 13, 14, 20, 23, 24, 28
36 (39.9, 25.1) 6, 13, 14, 20, 23, 24, 26, 28
37 (20.3, 18.2) 1, 2, 3, 4, 8, 11, 15, 19, 21, 27
38 (10, 20.5) 6, 7, 10, 13, 14, 20, 23, 24, 26, 28
39 (20.5, 21.4) 1, 2, 3, 4, 8, 11, 15, 19, 21, 25
40 (37.1, 28.6) 7, 10, 13, 14, 20, 23, 24, 26
41 (44.1, 16.1) 1, 15, 21
42 (41.1, 6) 9, 29
43 (43, 18.8) 5, 9, 12, 16, 18, 22
44 (45.4, 24.2) 9, 12, 16, 17, 18, 30
45 (36.2, 41.2) 5, 9, 17, 27, 29, 30
46 (27.5, 32.3) 12, 16, 17, 18, 29, 30
47 (47.8, 13.8) 22, 27, 29, 30
48 (8.9, 14.8) 5, 30
49 (6.8, 6.2) 5, 9, 12, 16, 17, 27, 30
50 (11.7, 35.8) 1, 2, 3, 4, 8, 11, 15, 19, 21, 25
of each session in the 50-node network.
1 21 4 4
2 5 26 7
3 19 20 6
4 33 6 10
5 37 10 9
6 23 11 2
7 25 46 3
8 42 43 9
9 44 27 8
10 47 30 1

d(9)
ten-session network. s(4)
40 d(6)
35 s(7) d(7)
d(1)
30 d(5)
s(3)
d(2)
25 d(4) s(9)
s(6)
20
s(5) d(8)
15
s(1) s(10)
10
d(3)
5 s(8)
s(2) d(10)
0
0 5 10 15 20 25 30 35 40 45 50

can be applied to any polynomial term of the form ni=1 (xi )ci in variables xi ,
where the ci -exponents are constant integers. Given such generality, RLT is a
powerful tool in deriving tight linear relaxations.
As a case study, we considered a throughput maximization problem in a
multi-hop CRN under the SINR model. We developed a mathematical formu-
lation for joint optimization of power control, scheduling, and flow routing. We
presented a solution procedure based on the branch-and-bound framework and
applied RLT in deriving tight linear relaxations for a product of variables. In
this case study, we also learned how to identify the core optimization space for
the underlying problem and how to exploit different physical interpretations of
the core variables in developing a solution.
6.8 Problems
6.1 For the first simple example of RLT in Section 6.1, we showed that a
nonlinear (bilinear) term x1 x2 can be represented by a single variable X{1,2}
147 6.8 Problems
and four linear constraints (6.1) (6.4). Show that X{1,2} = x1 x2 holds true
when either x1 or x2 equals to its corresponding lower or upper bound value.
6.2 For Example 6.1, please show the three linear constraints for represent-
ing the relationship X{2,2,2,2} = X{2,2}
2 , and the four linear constraints for rep-
resenting X{1,2,2,2,2,2} = X{1,2} X{2,2,2,2} and X{1,2,2,2,2,2,3} = X{1,2,2,2,2,2} x3 ,
respectively. Derive appropriate lower and upper bounds on variables as nec-
essary for this representation.
6.3 In the formulation of the multi-hop CRN problem, describe the set of
constraints associated with each layer. Which set of constraints couples multi-
ple layers together?
6.4 By introducing new variables tkm , we can replace (6.22) by (6).24 and
decrease the number of nonlinear terms. Analyze the number of nonlinear
terms before and after this substitution. What is the effect of this reformula-
tion?
6.5 Verify that (6.21), (6).24, and (6.25) imply (6.16).
6.6 For all the variables in the MINLP formulation for the multi-hop CRN
problem, what do we mean by core variables? What is the benefit of identi-
fying core variables in developing a solution?
6.7 Show how the values for the dependent variables tim , sijm , fij (l), and K
can be derived from the values for the core variables xijm and qijm .
6.8 When determining upper bounds for a subproblem, we need to perform
linear relaxation for two different nonlinear terms. Describe the technique we
used in linear relaxation of each nonlinear term. Can each technique in linear
relaxation be applied for both nonlinear terms?
6.9 In the local search, why do we initialize qijm at (qijm )L and increase it
upward? Discuss the case of initializing qijm at (qijm )U and decreasing it down-
ward.
6.10 When choosing a partitioning variable, why do we consider that the
x-variables are more important than the q-variables?
6.11 After partitioning a problem into two new subproblems on a xijm -
variable, how can the lower and upper bounds for the other core variables be
adjusted by using the first five constraints in Table 6.2?
CHAPTER
7 Linear approximation
The great pleasure in life is doing what people say you cannot do.
Walter Bagehot
7.1 Review of linear approximation for nonlinear terms
In contrast to LP, nonlinear optimization problems are usually hard and tedious
to solve. Therefore, one strategy for nonlinear programs is to replace all non-
linear terms by some linear approximation approach [11]. Depending on the
nature of the nonlinear program, the obtained LP problem may or may not
provide a feasible solution to the original nonlinear problem. In the case that
the LP solution is feasible, we may prove that it is a near-optimal solution
to the original problem. Even if it is infeasible, we may construct a feasible
solution via local search and further prove the constructed solution is near-
optimal.
In this chapter, we will explore such an approach to some nonlinear pro-
grams with special structures. This approach is also known as grid lineariza-
tion [88] or separable programming [11]: We consider the following problem:
Minimize f (x)
subject to gi (x) pi (1 i I )
hj (x) = qj (1 j J )
x 0,
where the objective function f (x) and constraint functions gi (x) are additively
separable, the constraint functions hj (x) are linear, and pi and qj are constants.
In other words, this implies that each f (x) and gi (x) can be expressed as a sum
of functions as follows:
148
149 7.1 Review of linear approximation for nonlinear terms
()
()
Figure 7.1 Piecewise linear

N
approximation. f (x) = fn (xn )
n=1

N
gi (x) = gin (xn ) pi (1 i I ).
n=1
A general approach to approximate a nonlinear function is to use a piecewise

linear function [11]. Let us consider a continuous function (), where
[a, b] (see Fig. 7.1). In order to define an approximating function (), we
divide the interval [a, b] into smaller intervals [k1 , k ], 1 k m, through
the grid points 0 , 1 , 2 , . . . , m , where 0 = a and m = b. Note that any
[k1 , k ] can be represented by
= k1 k1 + k k
k1 + k = 1
k1 , k 0 .
The original function () in the interval [k1 , k ] can be approximated

by
() = k1 (k1 ) + k (k ).
It is quite apparent that as the number of grid points increases, the accuracy of
the approximation improves.
In general, for any [a, b], we can represent it as

m
= k k (7.1)
k=0

m
k = 1, (7.2)
k=0
where only two continuous k1 and k can be positive (when falls in the
kth segment). All other l , l = k 1, k, must be zero. The original function
() is approximated by

m
() = k (k ) . (7.3)
k=0
To model the relationship that when falls in the kth segment, k1 0,

k 0, and l = 0 for l = k 1, k, we define binary variables zk , 1 k m,
to indicate whether falls within the kth segment [k1 , k ]. That is, zk = 1
if k1 < k and zk = 0 otherwise. Since can only fall in one of the m
segments, we have

m
zk = 1 . (7.4)
k=1
To model the relationship between k and zk , we note that when falls in

the kth segment (i.e., zk = 1), k1 and k may be positive, while all other
l (l = k 1, k) must be zero. That is, 0 > 0 only if z1 = 1; k > 0 only
if zk = 1 or zk+1 = 1, k = 1, . . . , m 1; and m > 0 only if zm = 1. These
relationships can be written as follows:
0 z1 (7.5)
k zk + zk+1 (1 k < m) (7.6)
m zm . (7.7)
Therefore, piecewise linear approximation can be formulated by constraints
(7.1)(7.7). We note here that there are other improved strategies to model such
piecewise linear functions, which afford partial convex hull representations
[147].
Using such piecewise linear approximations, we can derive a linear approx-
imating problem for the original problem by replacing each nonlinear function
fn (xn ) and gin (xn ) by its piecewise linear approximation. The function hj (x)
is already linear and no effort is needed to approximate it. Therefore, the new
problem is a linear mixed-integer program.
Suppose we solve the linear approximating problem and obtain its solution
. It has been shown that if suitable convexity requirements hold (i.e., both the
objective functions fn (xn ) and the constraint functions gin (xn ) are convex),
then even without the z-variables and the constraints (7.4)(7.7), the solution
is a feasible solution to the original problem [11]. Note that each function
hj (x) is linear. Thus, the above requirements hold if and only if the original
problem is convex. In this case, we can further analyze the performance of .
Since the accuracy of piecewise linear approximation is closely related with
the number of grid points used in approximating each function, we can expect
that the performance of improves as the number of grid points increases.
However, as the number of grid points increases, the number of variables in the
151 7.2 Case study: Renewable sensor networks with wireless energy transfer
linear approximating problem also increases, leading to a higher complexity.

Thus, by adjusting the number of grid points in the linear approximation, we
can make a tradeoff between accuracy in near-optimality and complexity.
For the case that the original problem is not convex, is likely to be infeasi-
ble to the original problem. But since the new problem is an approximation of
the original problem, is near-feasible, i.e., the violation of each constraint is
very small. Then we may construct a feasible solution based on by a local
search, which is a problem-specific procedure. In the following sections, we
will demonstrate the use of piecewise linear approximations and local search
to solve a nonconvex program for a wireless sensor network (WSN).
7.2 Case study: Renewable sensor networks with wireless

energy transfer
Wireless sensor networks (WSNs) today are mainly powered by batteries. Due
to limited energy storage capacity in a battery at each node, a WSN can only
remain operational for a limited amount of time. To prolong its lifetime, there
have been a flourish of research efforts in the last decade (see, e.g., [24], [163],
[50], [71], [151]). Despite these intensive efforts, lifetime remains a perfor-
mance bottleneck of a WSN and is perhaps one of the key factors that hinder
its wide-scale deployment.
Although energy-harvesting (or energy scavenging) techniques (see, e.g.,
[76], [77], [21, Chapter 9], [121]) have been proposed to extract energy from
the environment, their success remains limited in practice. This is because the
proper operations of any energy-harvesting technique are highly dependent on
the environment. Further, the size of an energy-harvesting device may pose
concern in deployment, particularly when the size of such a device is of much
larger physical scale than the sensor node it is attempting to power.
Quite unexpectedly, the recent breakthrough in the area of wireless energy
transfer technology, developed by Kurs et al. [85], has opened up a revolution-
ary paradigm for prolonging sensor network lifetime. Basically, Kurs et al.s
work showed that by exploiting a novel technique called magnetic resonant
coupling, wireless power transfer (i.e., the ability to transfer electric power
from one storage device to another without any plugs or wires) is both feasi-
ble and practical. In addition to wireless power transfer, they experimentally
showed that the source energy storage device does not need to be in contact
with the energy receiving device (e.g., a distance of 2 meters) for efficient
energy transfer. Moreover, wireless power transfer is immune to the neighbor-
ing environment and does not require a line of sight between the power charg-
ing and receiving nodes. Recent advances in this technology further show that
it can be made portable, with applications to palm size devices such as cell
phones [49].
The impact of wireless energy transfer on WSNs or other energy-constrained

wireless networks is immense. Instead of generating energy locally at a node
(as in the case of energy harvesting), we can bring clean electric energy effi-
ciently generated elsewhere to a sensor node periodically, and charge its battery
without the constraint of wires and plugs. As we can imagine, the applications
of wireless energy transfer are numerous. For example, wireless energy trans-
fer has already been applied to replenish battery energy in medical sensors and
implantable devices [178] in the healthcare industry.
Inspired by this new breakthrough in energy transfer technology, this
chapter re-examines the network lifetime paradigm for a WSN. We envision
employing a mobile vehicle carrying a power charging station to periodically
visit each sensor node and charge it wirelessly. This mobile wireless charging
vehicle (WCV) can either be manned by a human or be entirely autonomous.
In this chapter, we investigate the fundamental question of whether such a new
technology can be applied to remove the lifetime performance bottleneck of a
WSN. That is, through periodic wireless re-charge, we show that each sensor
node will always have an energy level above a minimum threshold so that the
WSN remains operational forever. Some of the highlights of this chapter are
as follows:
We introduce the concept of a renewable energy cycle where the remain-

ing energy level in a sensor nodes battery exhibits some periodicity over a
time cycle. We offer both necessary and sufficient conditions for a renewable
energy cycle and show that feasible solutions satisfying these conditions can
offer renewable energy cycles and, thus, unlimited sensor network lifetime.
We investigate an optimization problem, with the objective of maximizing
the ratio of the WCVs vacation time (time spent at its home station) over
the cycle time. In terms of achieving the maximum ratio, we prove that the
optimal traveling path for the WCV in each renewable cycle is the shortest
Hamiltonian cycle. We also derive several interesting properties associated
with an optimal solution, such as the optimal objective being independent of
traveling direction on the shortest Hamiltonian cycle and the existence of an
energy bottleneck node in the network.
Under the optimal traveling path, our optimization problem now only needs
to consider flow routing and the charging time for each sensor node. We for-
mulate an optimization problem for joint flow routing and charging schedule
for each sensor node. The problem is shown to be a nonlinear optimization
problem and is NP-hard in general. We show how to apply a piecewise lin-
ear approximation technique for each nonlinear term and obtain a tight linear
relaxation. Based on this linear relaxation, we obtain a feasible solution and
prove that it can achieve near-optimality for any desired level of accuracy.
The remainder of this chapter is organized as follows. In Section 7.3, we

review recent advances in wireless energy transfer technology. In Section 7.4,
153 7.3 Wireless energy transfer: a primer
we describe the scope of our problem for a renewable sensor network.

Section 7.5 introduces the concept of a renewable energy cycle and presents
some interesting properties. Section 7.6 shows that an optimal traveling path
should be along the shortest Hamiltonian path. In Section 7.7, we present our
problem formulation and a near-optimal solution. Section 7.8 shows how to
construct the initial transient cycle preceding the first renewable cycle. In Sec-
tion 7.9, we present numerical examples to demonstrate the properties of a
renewable wireless sensor network under our solution. Section 7.10 summa-
rizes this chapter.
7.3 Wireless energy transfer: a primer
Efforts at transferring power wirelessly can be dated back to the early 1900s
(long before wired electric power grid) when Nikola Tesla experimented with
large-scale wireless power distribution [156]. Due to its large electric fields,
which is undesirable for efficient energy transfer, Teslas invention was never
put into practice.
Since then, there was hardly any progress in wireless energy transfer for
many decades. In the early 1990s, the need for wireless power transfer re-
emerged when portable electronic devices became widely spread (see, e.g.,
[159]). The most well-known example is the electric toothbrush. However, due
to stringent requirements such as close contact, accurate alignment in charging
direction, and uninterrupted line of sight, most of the wireless power trans-
fer technologies at the time (based on inductive coupling) only found limited
applications.
Recently, wireless power transfer based on radio frequency (RF) between
850 MHz 950 MHz (with a center frequency of 915 MHz) has been explored
[125]. Under such radiative energy transfer technology, an RF transmitter
broadcasts radio waves in the 915 MHz ISM band and an RF receiver tunes
to the same frequency band to harvest radio power. However, it was found in
[93], [124], and [157] that a receiver operating under such radiative energy
transfer technology can only obtain about 45 mW power when it is 10 cm
away from the RF transmitter, with about 1% power transfer efficiency. A sim-
ilar experimental finding was reported in [66]. The technology is also sensitive
to obstructions between sources and devices, requires complicated tracking
mechanisms if relative positions change, and poses more stringent safety con-
cerns. Due to these issues, the potential of RF-based power transfer technology
is limited.
The foundation of this chapter is based on a recent breakthrough technol-
ogy by Kurs et al. [85], which was published in Science in 2007 and has
since caught worldwide attention. In [85], Kurs et al. experimentally demon-
strated that efficient nonradiative energy transfer was not only possible, but was
Figure 7.2 An example sensor Service station

network with a mobile WCV.
Sensor node
Mobile WCV
Base station
also practical. They used two magnetic resonant objects having the same reso-
nant frequency to exchange energy efficiently, while dissipating relatively little
energy in extraneous off-resonant objects. They showed that efficient power
transfer implemented in this way can be nearly omnidirectional, irrespective
of the environment and even none-line-of-sight (NLOS). The power transfer
efficiency, however, decreases with distance. A highlight of their experiment
was to fully power a 60-W light bulb from a distance of two meters away, with
about 40% power transfer efficiency.
Since the first demo by Kurs et al. in 2007, there has been some rapid
advance on wireless energy transfer, particularly in the area of making it
portable. In particular, Kurs et al. launched a start-up company [171] and in
2009 they developed and demonstrated wireless energy transfer for portable
devices such as cell phones [49]. Note that the source coil remains sizable,
but the device coil is already portable (corresponding to our WCV and sensor
node, respectively).
With the recent establishment of Wireless Power Consortium [172] to set
the international standard for interoperable wireless charging, it is expected
that wireless power transfer will revolutionize how energy is replenished in the
near future.
7.4 Problem description
We consider a set of sensor nodes N distributed over a two-dimensional area

(see Fig. 7.2). Each sensor node has a battery capacity of Emax and is fully
charged initially. Also, denote Emin as the minimum energy at a sensor node
battery (for it to be operational). For simplicity, we define network lifetime as
the time until the energy level of any sensor node in the network falls below
Emin [24; 133; 164]. Although a more general definition of network lifetime
155 7.4 Problem description
(e.g., [34]) is available, we decided to choose a simple network lifetime defini-

tion in this chapter, which is sufficient to show the potential of wireless energy
transfer in a sensor network.
Each sensor node i generates sensing data with a rate of Ri (in b/s), i N .
Within the sensor network, there is a fixed base station (B), which is the sink
node for all data generated by the sensor nodes. Multi-hop data routing can be
employed for forwarding data by the sensor nodes. Denote fij as the flow rate
from sensor node i to sensor node j and fiB as the flow rate from sensor node
i to the base station B, respectively. Then we have the following flow balance
constraint at each sensor node i:
k=i
j =i

fki + Ri = fij + fiB (i N ). (7.8)
kN j N
Each sensor node consumes energy for data transmission and reception.
Denote pi the energy consumption rate at sensor node i N . In this chap-
ter, we use the energy consumption model in Section 2.3.1. We have
k=i
j =i

pi = fki + Cij fij + CiB fiB (i N ) , (7.9)
kN j N
where is the energy consumption for receiving a unit of data, Cij (or
CiB ) is the energy consumption for transmitting a unit of data from node
i to node j (or the base station B). Further, Cij = 1 + 2 Dij , where Dij
is the distance between nodes i and j , 1 is a distance-independent con-
stant term, 2 is a coefficient of the distance-dependent term, and is the
k=i
path-loss index. In the model, kN fki is the energy consumption rate
j =i
for reception, and j N Cij fij + CiB fiB is the energy consumption rate for
transmission.
To re-charge the battery at each sensor node, a mobile wireless charging
vehicle (WCV) is employed in the network. The WCV starts from a service
station (S), and the traveling speed of the WCV is V (in m/s). When it arrives
at a sensor node, say i, it will spend an amount of time i to charge the sensor
nodes battery wirelessly via wireless power transfer [85]. Denote U as the
energy transfer rate of the WCV. After i , the WCV leaves node i and travels
to the next node on its path. We assume that the WCV has sufficient energy to
charge all sensor nodes in the network.
After the WCV visits all the sensor nodes in the network, it will return to
its service station to be serviced (e.g., replacing or re-charging its battery) and
get ready for the next trip. We call this resting period vacation time, denoted
as vac . After this vacation, the WCV will go out for its next trip.
Denote as the time for a trip cycle of the WCV. A number of important
questions need to be addressed for such a network. First and foremost, we
would ask whether it is possible for each sensor node never to run out of its
energy. If this is possible, then the sensor network will have unlimited lifetime
and will never cease to be operational. Second, if the answer to the first ques-
tion is positive, then is there an optimal plan (including traveling path, stop
schedule) such that some useful objective can be maximized or minimized?
For example, in this chapter, we would like to maximize the percentage of
time in a cycle that the WCV spends on its vacation (i.e., vac
), or equivalently,
to minimize the percentage of time that the WCV is out in the field.
Table 7.1 lists the notation used in this chapter.
7.5 Renewable cycle construction
In this section, we focus on the renewable cycle construction. We assume the

WCV starts from the service station, visits each sensor node once in a cycle
and ends at the service station (see Fig. 7.3). Further, we assume that the data
flow routing in the network is invariant with time, with both routing and flow
rates being part of our optimization problem.
The middle sawtooth graph (in dashed line) in Fig. 7.4 shows the energy
level of a sensor node i during the first two renewable cycles. Note that there
is an initialization cycle (marked in the grey area) before the first renewable
cycle. That initialization cycle will be constructed in Section 7.8 once we have
an optimal solution to the renewable cycles.
Denote P = (0 , 1 , . . . , N , 0 ) as the physical path traversed by the
WCV over a trip cycle, which starts from and ends at the service station (i.e.,
0 = S) and the ith node traversed by the WCV along path P is i , i N .
Denote D0 1 as the distance between the service station and the first sen-
sor node visited along P and Dk k+1 as the distance between the kth and
(k + 1)th sensor nodes, respectively. Denote ai as the arrival time of the WCV
at node i in the first renewable energy cycle. We have

i1 D
k k+1
i1
ai = + + k . (7.10)
V
k=0 k=1
Denote DP as the physical distance of path P and P = DP /V as the time

used for traveling over distance DP . Recall that vac is the vacation time the
WCV spends at its service station. Then the cycle time can be written as

= P + vac + i , (7.11)
iN

where iN i is the total amount of time the WCV spends near all the sensor
nodes in the network for wireless energy transfer.
157 7.5 Renewable cycle construction
Table 7.1 Notation.
ai Arrival time of the WCV at node i during the first renewable cycle
B Denotes the base station
Cij (or CiB ) Energy consumption for transmitting a unit of data from node i to
node j (or the base station B)
Dij (or DiS ) Distance from node i to node j (or the service station S)
DP The traveling distance during a renewable cycle under a traveling
path P
DTSP The minimal traveling distance during a renewable cycle
Emax Full battery capacity of sensor nodes
Emin Minimal energy reserve to keep a sensor node operational
Ei Starting energy of sensor node i in a renewable cycle with fully
re-charged battery
ei (t) Energy level of sensor node i at time t with fully re-charged battery
fij (or fiB ) Flow rate from sensor node i to node j (or the base station B)
h The reciprocal of
m The number of piecewise linear segments used to approximate
N The set of sensor nodes in the network
P The traveling path of the WCV
pi Energy consumption rate at sensor node i
Ri Data rate generated at sensor node i
S Denotes the service station
U Full charge rate of the WCV
ui Charge rate at sensor node i during the initial transient cycle
V Moving speed of the WCV
zi,k Binary variable indicating whether i falls within the kth segment
Path-loss index
1 A distance-independent constant term in energy consumption for
data transmission
2 A coefficient of the distance-dependent term in energy
consumption for data transmission
The power consumption coefficient for receiving data
Targeted approximation error, 0 < 1
i The ratio of the charging time at node i to the entire cycle time.
vac The ratio of the vacation time to the entire cycle time.
i An approximation of i2
i,k The weight of the grid point mk for
i
i The ith node traversed by the WCV along path P
Overall time spent during a renewable cycle
i Time span for the WCV to be present at the sensor node i for
re-charging battery
vac Vacation time during a renewable cycle
P Traveling time of the WCV during a renewable cycle under a
traveling path P
TSP Minimal traveling time of the WCV during a renewable cycle
Figure 7.3 A WCV periodically Mobile WCV

Service station
visits each sensor node and
charges its battery via wireless
Sensor node
energy transfer.
Base station
ei
Initial transient cycle First renewable cycle Second renewable cycle
Emax
Ei
Initial transient cycle
not shown here
Gi
(To be constructed later)
Emin t
0
0 ai ai + i 2 ai + ai + + i 3
Figure 7.4 The energy level of a

sensor node i during the first We formally define a renewable energy cycle as follows:
two renewable cycles (partially
re-charged v.s. fully re-charged).
Definition 7.1
The energy level of a sensor node i N exhibits a renewable energy cycle
if it meets the following two requirements: (i) it starts and ends with the
same energy level over a period of ; and (ii) it never falls below Emin .
During a renewable cycle, the amount of charged energy at a sensor node i

during i must be equal to the amount of energy consumed in the cycle (so as
to ensure the first requirement in Definition 7.1). That is
pi = i U (i N ) . (7.12)
Note that when the WCV visits a node i at time ai during a renewable energy
cycle, it does not have to re-charge the sensor nodes battery to Emax . This is
illustrated in Fig. 7.4, where Gi denotes the starting energy of sensor node i in
a renewable cycle and gi (t) denotes the energy level at time t (dashed sawtooth
graph). During a cycle [, 2 ], we see that the energy level has only two slopes:
(i) a slope of pi when the WCV is not at this node (i.e., noncharging period),
and (ii) a slope of (U pi ) when the WCV is charging this node at a rate of U
(i.e., charging period). It is clear that gi (ai ) gi (t) gi (ai + i ), i.e., node
is energy level is lowest at time ai and is highest at time ai + i .
Also shown in Fig. 7.4 is another renewable energy cycle (in solid sawtooth
graph) where the battery energy is charged to Emax during a WCVs visit.
For this energy curve, denote Ei as the starting energy of node i in a renewable
be an optimal
cycle and ei (t) as the energy level at time t, respectively. Let Full
solution with fully re-charged battery in each cycle that maximizes the ratio of
the WCVs vacation time over the cycle time. Let be an optimal solution,
where there is no requirement on whether or not a nodes battery is fully re-
charged. Naturally, the optimal objective obtained by Full is no more than the

optimal objective obtained by due to the additional requirement (battery is
. Surprisingly, the following lemma shows that is
fully re-charged) in Full Full
equally good as in terms of maximizing the ratio of the WCVs vacation
time over the cycle time. Thus, for our optimization problem, it is sufficient to
consider a solution with fully re-charged battery.
Lemma 7.1
can achieve the same maximum ratio of vacation time to cycle
Solution Full
time as that for solution .
Proof. Our proof has two parts. (i) First, we show that the maximal ratio of
vacation time to cycle time achieved by solution is greater than or equal to
that achieved by solution Full . (ii) Second, we show the converse is also true,
i.e., the maximal ratio of vacation time to cycle time achieved by solution Full

is also greater than or equal to that achieved by solution . If both (i) and (ii)
hold, then the lemma is proved.
Since Full is an optimal solution with the additional requirement (fully
re-charged battery in each cycle), the maximal ratio of vacation time to the
cycle time obtained by Full is no more than that obtained by . Thus, (i)
holds.
We now prove (ii). Instead of considering optimal solution , we will prove
that any ratio achieved by a feasible solution can also be achieved by a fea-
sible fully re-charged solution . If this is true, in the special case that =
is an optimal solution, we have that the maximal ratio achieved by can also
be achieved by a feasible fully re-charged solution. Therefore, (ii) will hold.
The proof is based on construction. Suppose = (P , ai , Gi , fij , fiB , ,
i , P , vac , pi ) is a feasible solution to our problem. We construct = (P , ai ,
Ei , fij , fiB , , i , P , vac , pi ) by letting P = P , ai = ai , Ei = Emax + pi
(ai + i ) U i , fij = fij , fiB = fiB , = , i = i , P = P , vac =

vac , and pi = pi . Note that under , the maximal energy level of node i occurs
at time (ai + i ), which is ei (ai + i ) = Ei + U i pi (ai + i ) =
[Emax + pi (ai + i ) U i ] + U i pi (ai + i ) = Emax . Thus,
is a fully re-charged solution. Moreover, it is clear that vac
= vac
since =
and vac = vac . Now, all we need to do is to verify that is a feasible renew-
able cycle.
To show that is feasible, we need to verify that meets constraints (7.8),
(7.9), and (7.11), as well as ei (t) Emin for i N , t . Further, to show
that is a renewable cycle, we need to verify that ei (k ) = Ei for k N.
We now verify each of these requirements. Since is a feasible solution, it
satisfies (7.8), (7.9), and (7.11). In , we have = , i = i , P = P , vac =
vac , fij = fij , fiB = fiB , and pi = pi . Then also satisfies (7.8), (7.9), and
(7.11).
We now show ei (t) Emin for i N , t . Since is a feasible solu-
tion, Emax gi (ai + i ) = Gi + U i pi (ai + i ). Thus, we have Emax
+ pi (ai + i ) U i Gi . Since Ei is set as Ei = Emax + pi (ai + i
) U i , we have Ei Gi . Moreover, since pi = pi , the energy at a node i
in , ei (t), is parallel to gi (t) in . Because of the parallelism and Ei Gi ,
we have ei (t) gi (t) for i N . Since is a feasible solution, we have
gi (t) Emin for i N . Thus, ei (t) gi (t) Emin for i N . Therefore,
is feasible.
Because of the parallelism and gi (k ) = Gi for k = 1, 2, . . ., we have
ei (k ) = Ei for k = 1, 2, . . .. Thus, is a feasible renewable cycle, and (ii)
holds.
Based on Lemma 7.1, we will only consider a renewable cycle where each
node is fully re-charged when it is visited by the WCV. Since the energy level
at node i is at its lowest at time ai , to ensure the second requirement in Defini-
tion 7.1 we must have ei (ai ) = Ei (ai )pi Emin . Since for a renewable
cycle,
Ei = ei (2 ) = ei (ai + i ) (2 ai i )pi = Emax (2 ai i )pi ,

(7.13)
we have ei (ai ) = Emax (2 ai i )pi (ai )pi = Emax ( i )

pi . Therefore,
Emax ( i ) pi Emin (i N ). (7.14)
To construct a renewable energy cycle, we need to consider the travel-

ing path P , the arrival time ai , the starting energy Ei , the flow rates fij
and fiB , time intervals , i , P , and vac , and power consumption pi . By
(7.10) and (7.13), ai and Ei are variables that depend on P , , and i .
Thus, ai and Ei can be excluded from a solution , with the result that
= (P , fij , fiB , , i , P , vac , pi ). Although more variables can be removed

from , we keep this representation for the sake of future discussion.
For a renewable energy cycle, we have the following lemma:
Lemma 7.2
A cycle is a renewable energy cycle if and only if constraints (7.11), (7.12),
and (7.14) are met for each sensor node i N .
Proof. The only if part of the lemma can be proved by showing that a renew-
able cycle meets (7.11), (7.12), and (7.14). This has already been shown in the
description of the renewable cycle.
We now prove the if part of the lemma, i.e., if (7.11), (7.12), and (7.14)
hold, then (i) and (ii) in Definition 7.1 will also hold, thus the cycle is a renew-
able energy cycle. Since (7.11) holds, the given cycle satisfies the time con-
straint. Constraint (7.12) ensures that the amount of energy charged to each
sensor node i during i is equal to the amount of energy consumed by sensor
node i in the cycle. So the energy level of each sensor node i at the end of the
cycle is the same as that at the beginning of the cycle. Therefore, requirement
(i) in Definition 7.1 is satisfied.
During the first renewable cycle, the lowest energy level at node i occurs at
time ai , which is
ei (ai ) = ei ( ) (ai )pi

= ei (2 ) (ai )pi
= Emax (2 ai i )pi (ai )pi
= Emax ( i )pi
Emin ,
where the second equality holds by requirement (i), which we just proved, the
third equality holds by (7.13), the fourth equality holds by ei (ai + i ) = Emax
in a fully re-charged solution, and the last inequality holds by (7.14). Since
the lowest energy level of node i occurs at time ai and is still no less than
Emin , requirement (ii) in Definition 7.1 is met. The proof of the if part of the
lemma is complete.
The following property shows that in an optimal solution, there exists at

least one bottleneck node in the network, where the energy level at this node
drops exactly to Emin upon the WCVs arrival.
Property 7.1
In an optimal solution, there exists at least one node in the network with its
battery energy dropping to Emin when the WCV arrives at this node.
Proof. The proof is based on contradiction (i.e., if this is not true, then we can
further increase the objective value, thus leading to contradiction).
Suppose there exists an optimal solution = (P , fij , fiB , , ,
i
, p ), where none of the nodes in the network has its energy
P , vac i
ever drops to Emin , i.e., ei (t) > Emin for all i N , t . Then we can
construct a new solution = (P , fij , fiB , , i , P , vac , pi ) by choosing
Emin
= miniN { E(max )p } 1 and letting P = P , fij = fij , fiB = fiB , =

i i

(1 + ) , i = (1 + )i , P = P , vac = vac + ( iN i ), and
pi = pi .
Now we show > 0. Since ei (t) > Emin for all i N , t , we have
ei (ai ) = Emax ( i )pi > Emin for all i N , i.e., miniN

{Emax ( i )pi } > Emin . It follows that Emax maxiN {(

Emin Emax Emin
i )pi } > Emin , or max Emax {( )p } > 1. Thus, = miniN { ( )p }
iN i i i i
Emin
1 = max Emax 1 > 0.
iN {( i )pi }
The feasibility of can be verified similarly to that in the proof of
Lemma 7.1.
We now show that this new feasible solution can offera better (increased)
i
objective value. By (7.11), we have vac

= 1 P iN
. Since =

(1 + ) , i = (1 + )i , P = P , it follows that vac

=1 P
(1+ )

iN (1+ )i
i
vac vac
vac
(1+ ) > 1 P iN = ,
i.e.,
> . This contradicts
the assumption that is an optimal solution.
7.6 Optimal traveling path
In this section, we show that the WCV must move along the shortest Hamil-
tonian cycle in an optimal solution. This is formally stated in the following
theorem:
Theorem 7.1
In an optimal solution with the maximal vac
, the WCV must move along the
shortest Hamiltonian cycle that crosses all the sensor nodes and the service
station.
Proof. Theorem 7.1 can be proved by contradiction. That is, if there is an opti-
mal solution = (P , fij , fiB , , , , , p ), where the WCV does
i P vac i
not move along the shortest Hamiltonian cycle, then we can construct a new
solution = (P , fij , fiB , , i , P , vac , pi ), with the WCV moving along the
shortest Hamiltonian cycle and with an improved objective.
By assumption, P in does not follow the shortest Hamiltonian cycle.
The new solution is constructed as follows. Let P follows the shortest hamil-
163 7.6 Optimal traveling path
tonian cycle (by either direction), fij = fij , fiB = fiB

, = , = ,
i i
pi = pi , P is the traveling time spent on path P , and

vac = vac + P P . (7.15)
Now we show the constructed solution is feasible. To verify the fea-
sibility, we need to show that satisfies the flow conservation constraint
(7.8), the time constraint (7.11), and the energy constraints (7.12), and
(7.14). Since is a feasible solution, it satisfies (7.8), (7.11), (7.12), and
(7.14). Since we have fij = fij , fiB = fiB , = , = and p = p
i i i i
in , the constraints (7.8), (7.12), and (7.14) also hold by . To show

that also satisfies (7.11), we have P + iN i + vac = P + iN i +

+
vac P
P = iN i + vac +
P
= = , where the first equal-
ity follows from (7.15), the second equality follows by the feasibility of and
(7.11), and the third equality follows by = during construction.
To show vac / > vac / , recall that P follows the shortest Hamiltonian

cycle while P does not, i.e., the traveling distance DP > DP . Therefore,
the traveling time P > P . Then by (7.15), vac = vac + , or
P > vac
P
/ = / . But / > / contradicts the assumption
vac / > vac vac vac vac

that is optimal. This completes the proof.
Theorem 7.1 says that the WCV should move along the shortest Hamiltonian
cycle, which can be obtained by solving the well-known Traveling Salesman
Problem (TSP) (see, e.g., [28; 119]). Denote DTSP as the traveling distance in
the shortest Hamiltonian cycle and let TSP = DTSP /V . Then with the optimal
traveling path, (7.11) becomes

TSP + vac + i = , (7.16)
iN
and the solution for a renewable cycle becomes = (PTSP , fij , fiB , , i ,
TSP , vac , pi ). Since the optimal traveling path is determined, the solution can
be simplified as = (fij , fiB , , i , vac , pi ).
We note that the shortest Hamiltonian cycle may not be unique. Since any
shortest Hamiltonian cycle has the same total path distance and traveling time
TSP , the selection of a particular shortest Hamiltonian cycle does not affect
constraint (7.16), and yields the same optimal objective. This insight is for-
mally stated in the following corollary:
Corollary 7.1
Any shortest Hamiltonian cycle can achieve the same optimal objective.
We also note that to travel the shortest Hamiltonian cycle, there are two
(opposite) outgoing directions for the WCV to start from its home service sta-
tion. Since the proof of Theorem 7.1 is independent of the starting direction for
the WCV, either direction will yield an optimal solution with the same objec-
tive value, although some variables in each optimal solution will have different
values. We have the following corollary:
Corollary 7.2
The WCV can follow either direction to traverse the shortest Hamiltonian
cycle, both of which will achieve the same optimal objective. There exist
two optimal solutions corresponding to the two opposite directions, with
identical values of fij , fiB , , i , TSP , vac , pi , but different values of ai
(by (7.10)) and Ei (by (7.13)) due to difference in their respective renewable
cycles, where i, j N , i = j .
7.7 Problem formulation and solution

7.7.1 Mathematical formulation
Summarizing the objective and all the constraints in Sections 7.4, 7.5, and 7.6,
our problem can be formulated as follows:
OPT
vac
Maximize

j =i
k=i

subject to fij + fiB fki = Ri (i N )
j N kN
k=i
j =i

fki + Cij fij
kN j N
+ CiB fiB pi = 0 (i N ) (7.17)

i vac = TSP (7.18)
iN
p i U i = 0 (i N ) (7.19)
( i ) pi Emax Emin (i N ) (7.20)
fij , fiB , i , , vac , pi 0 (i, j N , i = j ).
In this problem, flow rates fij and fiB , time intervals , i , and vac , and power
consumption pi are optimization variables, and Ri , , Cij , CiB , U , Emax , Emin ,
and TSP are constants. This problem has both nonlinear objective ( vac ) and
nonlinear terms (pi and i pi ) in constraints (7.19) and (7.20).
Note that there are two possible outcomes for optimization problem OPT:
either an optimal solution exists or OPT is infeasible. There are several sce-
narios where the latter outcome may occur, e.g., (i) the energy charging rate of
165 7.7 Problem formulation and solution
WCV is too small or the energy consumption rate of a node is too large; (ii)
the time interval between WCVs visits at any node is too large. As a result,
some constraints in Problem OPT will not hold. These are physical limitations
for a WCV to achieve a renewable network lifetime for a WSN.
We note that in the above formulation, only constant TSP is related to the
shortest Hamiltonian cycle. Since this value does not depend on the traveling
direction along the Hamiltonian cycle, an optimal solution to Problem OPT
will work for either direction and yields a different renewable cycle for each
direction.
7.7.2 Reformulation
We first use the change-of-variable technique to simplify the formulation. For
the nonlinear objective vac
, we define
vac
vac = . (7.21)

For (7.18), we divide both sides by and have TSP 1 + vac + iN i = 1.
To remove the nonlinear terms 1 and i in the above equation, we define
i
i = (i N ) (7.22)

1
h= . (7.23)

Then (7.18) is reformulated as TSP h + vac + iN i = 1, or equivalently,

1 iN i vac
h= . (7.24)
TSP
Similarly, (7.19) and (7.20) can be reformulated (by dividing both sides
by ) as
pi = U i (i N ), (7.25)
(1 i ) pi (Emax Emin ) h (i N ) (7.26)
By (7.24) and (7.25),

constraint (7.26) can be rewritten as (1 i ) U i
1 kN k vac
(Emax Emin ) , or
TSP
U TSP
vac 1 k i (1 i ) (i N ) .
Emax Emin
kN
By (7.25), constraint (7.17) can be rewritten as

k=i
j =i

fki + Cij fij + CiB fiB U i = 0 (i N ).
kN j N
Algorithm 7.1 Algorithm to problem OPT

Once we solve problem OPT-R, we can obtain the solution to problem OPT
(i.e., calculate the values for , i , vac , and pi ) as follows: h by (7.24), by
(7.23), i by (7.22), vac by (7.21), and pi by (7.25).
Now problem OPT is reformulated as follows:
OPT-R
Maximizevac
j =i
k=i

subject to fij +fiB fki = Ri (i N )
j N kN
k=i
j =i

fki + Cij fij + CiB fiB U i = 0 (i N ) (7.27)
kN j N
U TSP
vac 1 k i (1 i ) (i N ) (7.28)
Emax Emin
kN
fij , fiB 0, 0 i , vac 1 (i, j N , i = j ).
In this problem, fij , fiB , i , and vac are optimization variables, and Ri , ,
Cij , CiB , U , Emax , Emin , and TSP are constants. Algorithm 7.1 shows how to
obtain a solution to problem OPT once we obtain a solution to problem OPT-R.
After reformulation, the objective function and the constraints become lin-
ear except (7.28), where we have a second order i2 term, with 0 i 1.
In the next section, we present an efficient linear approximation technique
to approximate this second order nonlinear term (with performance guaran-
tee). This is the main technique that we wish to show in this chapter. Sub-
sequently, we develop an efficient near-optimal solution to our optimization
problem.
Remark 7.1
In our optimization problem, data routing and charging time are closely
coupled. We may want to de-couple routing from the charging problem
and require certain energy-efficient routing, e.g., the minimum energy
routing.1 However, minimum energy routing cannot guarantee optimal-
ity. This is because, to maximize vac , by (7.28), we need to minimize
U TSP
maxiN { kN k + Emax E min
i (1 i )}, i.e., minimize kN k +
U
maxiN { Emax E
TSP
i (1 i )}. But under minimum energy routing, we
min
k=i j =i
can only guarantee that iN ( kN fki + j N Cij fij + CiB fiB ) is
k=i
minimized. By the relationship in (7.27), minimizing iN ( kN fki +
j =i
j N Cij fij + CiB fiB ) is equivalent to minimizing iN i , which
U TSP
is only part of kN k + maxiN { Emax E min
i (1 i )}. Therefore,
minimum energy routing cannot guarantee the optimality of our problem.
This insight will be confirmed by our numerical examples in Section 7.9.
7.7.3 A near-optimal solution

Roadmap Our roadmap to solve problem OPT is as follows. First, we
employ the piecewise linear approximation technique (see Section 7.1) for the
quadratic terms (i2 ) in problem OPT-R. This approximation relaxes the cor-
responding nonlinear constraints into linear constraints, which allows for the
problem to be solved by a solver such as CPLEX [31]. Based on the solu-
tion from CPLEX, we construct a feasible solution to problem OPT. In Sec-
tion 7.7.4, we prove the near-optimality of this feasible solution.
Piecewise linear approximation for i2 Note that the only nonlinear terms
in the formulation are i2 , i N . Further, i lies in the interval [0, 1], which
is small. This motivates us to employ the piecewise linear approximation tech-
nique for the quadratic terms i2 .
Based on the discussion in Section 7.1, for a piecewise linear approximation
with m grid points (see Fig. 7.5), we have the following constraints for i2 :

m
zik = 1 (7.29)
k=1

m
k
i = ik (7.30)
m
k=0

m
ik = 1 (7.31)
k=0

m
k2
i = ik (7.32)
m2
k=0
i0 zi1 (7.33)
ik zik + zi,k+1 (1 k < m) (7.34)
im zim , (7.35)
where zik is a binary variable indicating whether i falls within the kth segment
and ik [0, 1] is a weight associated with grid point mk . The setting of m will
determine the level of accuracy and will be studied in Section 7.7.4.
Since y = x 2 is a convex function, the piecewise linear approximation curve
(i , i ) lies above the curve (i , i2 ), 0 i 1. Thus, we have i i2 (see
Fig. 7.5). The following lemma characterizes the approximation error i i2
as a function of m.
Figure 7.5 An illustration of

Y
piecewise linear approximation
1
(with m = 4) for the curve
(i , i2 ), 0 i 1.
i
2i
1 1 3
0 4 i 2 4 1 X
Lemma 7.3
1
i i2 for i N .
4m2
Proof. Assume i falls in the kth segment, k = 1, 2, . . . , m. Then we have

% & % &
(k 1)2 k2 k1 k 2
i i2 = i,k1 + i,k i,k1 + i,k
m2 m2 m m
(k1) 2 k 2
= (i,k1 2i,k1 ) + (i,k 2i,k ) 2
m2 m
k1 k
2i,k1 i,k
m m
(k 1)2 k2
= (1 i,k ) (1 i,k )2 + ( i,k 2
i,k )
m2 m2
k1 k
2(1 i,k )i,k
% m m &
(k 1) 2 k2 k1 k
= (i,k 2i,k ) + 2
m2 m2 m m
% &2
k k1
= (i,k 2i,k )
m m
1
= (i,k 2i,k ) 2
m
1
,
4m2
where the first equality holds by (7.30), (7.32), and the fact that il = 0 for l =
k 1, k, the third equality holds by (7.31) and il = 0 for l = k 1, k, and the
last inequality holds by i,k 2i,k 14 when 0 i,k 1. This completes the
proof.
Relaxed linear formulation By replacing i2 with i in (7.28), we have
U TSP
vac 1 k (i i ) (i N ). (7.36)
Emax Emin
kN
By adding new constraints (7.29)(7.35), we obtain the following linear

relaxed formulation:
OPT-L
Maximize vac
j =i
k=i

subject to fij + fiB fki = Ri (i N )
j N kN
k=i
j =i

fki + Cij fij + CiB fiB
kN j N
U i = 0 (i N )
U TSP
vac 1 k (i i ) (i N )
Emax Emin
kN

m
zik = 1 (i N )
k=1

m
k
i ik = 0 (i N )
m
k=0

m
ik = 1 (i N )
k=0

m
k2
i ik = 0 (i N )
m2
k=0
i0 zi1 0 (i N )
ik zik zi,k+1 0 (i N , 1 k < m)
im zim 0 (i N )
fij , fiB 0, 0 i , vac , i 1 (i, j N , i = j )
zik {0, 1} (i N , 1 k m)
0 ik 1 (i N , 0 k m),
where fij , fiB , i , vac , zik , ik , and i are variables, Ri , , Cij , CiB , U , Emax ,
Emin , and TSP are constants. The new formulation can be solved by a solver
such as CPLEX [31].
Construction of a feasible near-optimal solution The solution to problem
OPT-L is likely to be infeasible to problem OPT-R (and problem OPT). But
based on this solution, we can construct a feasible solution to problem OPT.
Suppose = (fij , fiB , i , vac , zik , ik , i ) is the solution to problem
OPT-L. By observing (fij , fiB , i , vac ), we find that it satisfies all con-
straints to problem OPT-R except (7.28). To construct a feasible solution =
(fij , fiB , i , vac ) to problem OPT-R, we let fij = fij , fiB = fiB , i = i .
For vac , in order to satisfy (7.28), we define
) ,
U TSP
vac = min 1 k i (1 i ) .
iN Emax Emin
kN
It is easy to verify that this newly constructed solution satisfies all the con-
straints for problem OPT-R. Once we have this solution to problem OPT-R, we
can easily find a solution to problem OPT via Algorithm 7.1.
7.7.4 Proof of near-optimality

In this section, we quantify the performance gap between the optimal objective
(unknown, denoted as vac ) and the objective (denoted as ) obtained by the
vac
feasible solution that we derived in the last section. Naturally, we expect
such a performance gap to be a function of m, i.e., the number of segments
that we use in the piecewise linear approximation. This result will be stated in
Lemma 7.4. Based on this result, we can obtain an important inverse result (in

Theorem 7.2), which shows how to set m such that vac vac for a given
target performance gap (0 < 1).
Lemma 7.4

For the feasible solution with objective value vac , we have vac vac
U TSP
4(Emax Emin ) 1
m2
.
To prove Lemma 7.4, we need two intermediate results for vac and vac ,
which are stated in Lemmas 7.5 and 7.6, respectively.
Lemma 7.5
For the optimal solution to problem OPT-L, we have
U TSP
vac = 1 k (max max ) ,
Emax Emin
kN
Y
1
max
2
max
0 1 max 12 3
4
1 X
4
2
Figure 7.6 An illustration of max and its approximation max .
where
max max{i }
iN
2
and max is the piecewise linear approximation of max (see Fig. 7.6).
U TSP
Proof. It is easy to see that in , vac = miniN {1 kN k Emax Emin
(i i )}. Thus, to prove this lemma, it is sufficient to show maxiN {i
i } = max max , i.e., i i max max for each i N .
We consider the following two cases:
Case 1: 0 < max 12 : Based on the definition of max , we have 0 < i
max 12 for each i N . We now show that i i is a nondecreasing func-
tion when 0 < i 12 . Then it follows that i i max max .
m i < m for a particular
Suppose i falls within the lth segment, i.e., l1 l
l. We have
i,l1 = l m i and i,l = m i (l 1) .
Then we can express i i as a function of i .

% &
(l 1)2 l2
i i = i i,l1 + i,l 2
m2 m
(l 1)2 l2
= i (l m i ) + (m i (l 1))
m2 m2
(l 2 (l 1)2 ) mi l(l 1)
= i
% & m2
2l 1 l(l 1)
= 1 i + . (7.37)
m m2
When 0 < i 12 , we have l m+1 2 if m is odd and l 2 if m is even. Then
m
1 m 0 for both cases. Thus, when 0 < i 2 and i falls within the lth
2l1 1
segment, i i is a nondecreasing function. Since the nondecreasing property

holds for all segments within (0, 12 ], we know that i i is a nondecreasing
function when 0 < i 12 .
Case 2: 12 < max 1: Suppose j = max . Let j = 1 j and denote
j the approximation of j2 . By (7.36), we have i 1 j
! "
k=i,k=j U TSP
kN k + vac + Emax Emin ( i i ) < 1 j = j < 12 for each
i N . Since i i is a nondecreasing function for 0 < i 12 , we have
i i j j .
We now show j j = j j . Suppose j falls within the lth segment,
m j < m for a particular l. Then j falls within the (m l + 1)th
i.e., l1 l
segment, i.e., m j < m . Then by (7.37), we have

ml
ml+1
j j
% &
2(m l + 1) 1 (m l + 1)(m l + 1 1)
= 1 j +
m m2
m + 2l 1 (m l + 1)(m l)
= (1 j ) +
m m2
% &
2l 1 (m +2lmm) + (m2 2lm+l 2 +ml)
2
= 1 j +
m m2
% &
2l 1 l2 l
= 1 j +
m m2
= j j .
Therefore, we have i i j j = j j = max max .

Combining both cases, we have i i max max for each i N . This
completes the proof.
Lemma 7.6
For the constructed solution to problem OPT-R, we have
U TSP
vac = 1 k max (1 max ).
Emax Emin
kN
Proof. To prove the lemma, it is sufficient to show maxiN {i (1 i )} =

max (1 max ), i.e, i (1 i ) max (1 max ) for each i N .
We consider the following two cases:

Case 1: 0 < max 12 : We have 0 < i max 12 for each i N . Since
x(1 x) is an increasing function for 0 < x 12 , we have i (1 i )
max (1 max ).
Case 2: 12 < max < 1: Suppose j = max . By (7.36), we have i 1 j
! U TSP
"

kN ,k=i,k=j k + vac + Emax Emin ( i i ) < 1 j < 12 for each
i N . Since 0 < i < 1 j < 12 and x(1 x) is an increasing

function for
0 < x < 2 , we have i (1 i ) (1 j ) 1 (1 j ) = j (1 j ) =
1
max (1 max ).
Combining both cases, we have i (1 i ) max (1 max ). This completes
the proof.
Based on Lemmas 7.5 and 7.6, we can prove Lemma 7.4.
Proof of Lemma 7.4. Denote vac as the objective value obtained by solution
to the relaxed linear problem OPT-L. Since problem OPT-L is a relaxation
, i.e., . Therefore,
of problem OPT-R, vac is an upper bound of vac vac vac

vac vac vac vac
- .
U TSP
= 1 k (max max )
Emax Emin
kN
- .
U TSP
1 k max (1 max )
Emax Emin
kN
U TSP
= (max max
2
)
Emax Emin
U TSP 1
,
4(Emax Emin ) m2
where the second equality holds by Lemmas 7.5 and 7.6, the fourth inequality
holds by Lemma 7.3. This completes the proof.
Based on Lemma 7.4, the following theorem shows how to set m such that

vac vac for a given target performance gap (0 < 1).
Theorem 7.2 /0 1
U TSP
For a given , 0 < 1, if m = 4(Emax Emin ) , then we have vac
vac .

Proof. Lemma 7.4 shows that the performance gap /0 is vac vac 1
U TSP U TSP
4(Emax Emin ) m2 . Therefore, if we set m =
1
4(Emax Emin )
0
U TSP
4(Emax Emin ) , then we have
Algorithm 7.2 Summary of the procedure of how to construct a near-optimal

solution.
Construction of a near-optimal solution
1. Given a target
/0 performance1gap .
U TSP
2. Let m = 4(Emax Emin ) .
3. Solve problem OPT-L with m segments by CPLEX, and obtain its solution
= (fij , fiB , i , vac , zik , i,k , i ).
4. Construct a feasible solution = (fij , fiB , i , vac ) for problem OPT-R by

letting fij = fij , fiB = fiB , i = i and vac = miniN {1 kN k
U
Emax E
TSP
min
i (1 i )}.
5. Obtain a near-optimal solution (fij , fiB , , i , vac , pi ) to problem OPT by
Algorithm 7.1.
U TSP 1
vac vac
4(Emax Emin ) m2
U TSP 4(Emax Emin )

4(Emax Emin ) U TSP
=.
This completes the proof.
With Theorem 7.2, we display the complete solution procedure on how to

obtain a near-optimal solution to problem OPT in Fig. 7.7.
7.8 Construction of initial transient cycle
In Section 7.5, we skipped discussion on how to construct the initial tran-

sient cycle before the first renewable cycle. Now with the optimal traveling
path P (the shortest Hamiltonian cycle) and the feasible near-optimal solution
(fij , fiB , , i , vac , pi ) obtained in Section 7.7, we are ready to construct the
initial transient cycle.
Unlike a renewable energy cycle at node i, which starts and ends with the
same energy level Ei , the initial transient starts with Emax and ends with Ei .
Specifically, the initial transient cycle must meet the following criterion:
Criterion 7.1
At each node i N , its initial transient cycle must meet the following criteria:
(i) ei (0) = Emax and ei ( ) = Ei ; and (ii) ei (t) Emin for t [0, ].
175 7.8 Construction of initial transient cycle
ei
Initial transient cycle First renewable cycle
Emax
Ei
Emin
t
0 ai ai + i ai ai + i 2
Figure 7.7 Illustration of energy

behavior for the initial transient We now construct an initial cycle to meet the above criterion. First, we need
cycle and how it connects the to calculate Ei (i N ). From (7.13), we have Ei = Emax (2 ai i )pi ,
first renewable cycle. where ai can be obtained by (7.10).
For a solution = (P , fij , fiB , , i , P , vac , pi , U ) corresponding
to a renewable energy cycle for t , we construct = (P , fij ,
fiB , , i , P , vac , pi , ui ) for the initial transient cycle for t [0, ] by
letting P = P , fij = fij , fiB = fiB , = , i = i , = P , vac = vac ,
P
pi = pi , and
pi ai
ui = + pi , (7.38)
i
where ui is the charging rate at node i during the initial transient cycle and
ai is the arrival time of the WCV at node i in the initial transient cycle (see
Fig. 7.7).
Now we need to show ui U where U is the full charging rate. First, we
have

i1 D
k k+1
i1
i1 D
k k+1
i1
ai = + k = + k = ai , (7.39)
V V
k=0 k=1 k=0 k=1
where the second equality holds by P = P and i = i , and the last equality
follows from (7.10). Further, by (7.12), we have U i = pi = (2 )
pi (ai + i ) pi . It follows that
(ai ) pi (U pi ) i . (7.40)
Then, we have
pi ai pi (ai ) (U pi ) i
ui = + pi = + pi + pi = U ,
i i i
where the first equality follows from (7.38), the second equality follows from
(7.39), and the third inequality follows from (7.40).
For the newly constructed , we have the following theorem:
Theorem 7.3
The constructed is a feasible transient cycle.
Proof. To prove that is a feasible initial transient cycle, we need to show that
the newly constructed satisfies Criterion 7.1. By our assumption, ei (0) =
Emax . At time (ai + i ), we have
ei (ai + i ) = ei (0) pi ai + (ui pi ) i

= Emax pi (ai ) + (ui pi ) i
= Emax pi (ai + i ) + ui i
= Emax , (7.41)
where the second equality follows since ei (0) = Emax , pi = pi , ai = ai ,

and i = i , the last equality follows from (7.38) and (7.39). Therefore, the
battery at node i is full when the WCV leaves it at (ai + i ). At time , we
have

ei ( ) = ei (ai + i ) pi (ai + i )
= Emax pi [ (ai + i )]
= Emax pi [2 (ai + i )]
= ei (2 )
= Ei , (7.42)
where the second equality follows from (7.41), and the fourth equality follows
from (7.13). Therefore, Criterion 7.1(i) is met.
To show ei (t) Emin for t [0, ], it is sufficient to show that ei (ai )
Emin and ei ( ) Emin , since these two time instances are the local mini-
mum for ei (t) during t [0, ]. We have ei (ai ) = ei (0) pi ai = Emax
pi (ai ) Ei pi (ai ) = ei (ai ) Emin . Also, by (7.42), ei ( ) =
Ei Emin . Hence, ei (t) Emin , for t [0, ].
In summary, meets all the criteria of a feasible initial transient cycle. This
In this section, we present some numerical examples to demonstrate how our

solution can produce a renewable WSN and some interesting properties with
such a network.
7.9.1 Simulation settings

We consider two randomly generated WSNs consisting of 50 and 100 nodes,
respectively. The sensor nodes are deployed over a square area of 1 km
1 km. The data rate (i.e., Ri , i N ) from each node is randomly generated
within [1, 10] kb/s. The power consumption coefficients are 1 = 50 nJ/b,
2 = 0.0013 pJ/(b m4 ), = 4, and = 50 nJ/b [67]. The base station is
assumed to be located at (500, 500) (in m) and the home service station for the
WCV is assumed to be at the origin. The traveling speed of the WCV is V = 5
m/s.
For the battery at a sensor node, we choose a regular NiMH battery and its
nominal cell voltage and the quantity of electricity is 1.2 V/2.5 Ah. We have
Emax = 1.2 V 2.5 A 3600 sec = 10.8 KJ [94, Chapter 1]. We let Emin =
0.05 Emax = 540 J. We assume the wireless energy transfer rate U = 5 W,
which is well within the feasible range [85].
We set the target = 0.01 for the numerical results, i.e., our solution has an
error no more than 0.01.
7.9.2 Results
50-node network We first present complete results for the 50-node net-
work. Table 7.2 gives the location of each node and its data rate for a 50-
node network. The shortest Hamiltonian cycle is found by using the Concorde
solver [28] and is shown in Fig. 7.8. For this optimal cycle, DTSP = 5821 m
and TSP = 1164.2 s. For the target = 0.01, by Theorem 7.2, we have
23 4 23 4
U TSP 5 1164.2
m= = =4,
4(Emax Emin ) 4 0.01 (10800 540)
which is a small number. In our solution, the cycle time = 30.73 hours, the
vacation time vac = 26.82 hours, and the objective vac = 87.27%.
In Corollary 7.2, the WCV can follow either direction of the shortest Hamil-
tonian cycle while achieving the same objective value vac = 87.27%. Com-
paring the two solutions, the values for fij , fiB , , i , TSP , vac are identical,
while the values of ai and Ei are different. This finding can be verified by
our simulation results in Table 7.3 (counter clockwise direction) and Table 7.4
(clockwise direction). As an example, Figs. 7.9(a) and (b) show the energy
Table 7.2 Location and data rate Ri for each node in a 50-node network.
Node index Location (m) Ri (kb/s) Node index Location (m) Ri (kb/s)
1 (815, 276) 1 26 (758, 350) 9

2 (906, 680) 8 27 (743, 197) 1
3 (127, 655) 4 28 (392, 251) 4
4 (913, 163) 6 29 (655, 616) 4
5 (632, 119) 3 30 (171, 473) 9
6 (98, 498) 7 31 (706, 352) 5
7 (278, 960) 3 32 (32, 831) 10
8 (547, 340) 7 33 (277, 585) 1
9 (958, 585) 6 34 (46, 550) 3
10 (965, 224) 8 35 (97, 917) 2
11 (158, 751) 5 36 (823, 286) 2
12 (971, 255) 1 37 (695, 757) 8
13 (957, 506) 4 38 (317, 754) 6
14 (485, 699) 10 39 (950, 380) 6
15 (800, 891) 2 40 (34, 568) 2
16 (142, 959) 9 41 (439, 76) 8
17 (422, 547) 5 42 (382, 54) 7
18 (916, 139) 10 43 (766, 531) 4
19 (792, 149) 1 44 (795, 779) 6
20 (959, 258) 5 45 (187, 934) 6
21 (656, 841) 3 46 (490, 130) 1
22 (36, 254) 10 47 (446, 569) 3
23 (849, 814) 1 48 (646, 469) 2
24 (934, 244) 8 49 (709, 12) 2
25 (679, 929) 8 50 (755, 337) 3
cycle behavior of a sensor node (the 32nd node) under the two opposite travel-
ing directions, respectively.
By Property 7.1, we find that there exists an energy bottleneck node in the
network with its energy dropping to Emin during a renewable energy cycle.
This property is confirmed in our numerical results. This bottleneck node is
the 48th node, whose energy behavior is shown in Fig. 7.10.
In Section 7.7.2, we showed that minimum energy routing may not be opti-
mal for our problem (see Remark 7.1). This point is confirmed by our numeri-
cal results. In Fig. 7.11, we show that data routing in our solution differs from
the minimum energy routing for the 50-node network.
100-node Network Table 7.5 gives the location of each node and its data rate
for a 100-node network. The shortest hamiltonian cycle is shown in Fig. 7.12.
For this optimal cycle, DTSP = 7687 m and TSP = 1537.4 s. For the target
= 0.01, by Theorem 7.2, we have
23 4 23 4
U TSP 5 1537.4
m= = = 5.
4(Emax Emin ) 4 0.01 (10800 540)
179 7.11 Problems
Figure 7.8 An optimal traveling Y(m) Mobile WCV

path for the WCV for the
1000
50-node sensor network, Sensor node
assuming traveling direction is
counter clockwise.
500
Base station
Service station
X(m)
0
0 500 1000
The solution for the 100-node network includes the cycle time = 58.52 hours,
the vacation time vac = 50.30 hours, and the objective vac = 85.95%. Addi-
tional results are shown in Table 7.6.
7.10 Chapter summary
Linear approximation is a powerful approach to tackle certain nonlinear opti-

mization problems. In this chapter, we showed how such an approach could
be employed to solve a nonlinear programming problem in a WSN. In addi-
tion to the linear approximation technique, the problem in the case study is
interesting on its own, and demonstrates how the so-called wireless energy
transfer technology can be employed to address network lifetime problems in
a WSN.
7.11 Problems
7.1 Let us consider a continuous function () = 2 , where [0, 1] (see

Fig. 7.13). An approximating function () is defined by dividing the inter-
val [0, 1] into four smaller intervals [k1 , k ], 1 k 4, through the grid
points 0 = 0, 1 = 14 , 2 = 12 , 3 = 34 , 4 = 1. For any [0, 1], express
its approximation () using the piecewise linear approximation technique
Table 7.3 The case of counter clockwise traveling direction: Node visited along the
path, arrival time at each node, starting energy of each node in a renewable cycle,
and charging time at each node for the 50-node network.
Node visited Node visited

along ai (s) Ei (J) i (s) along ai (s) Ei (J) i (s)
the Path the Path
42 110702 10747 11 2 117778 10611 41

41 110725 10613 37 44 117848 10605 42
46 110777 9282 305 23 117903 10793 2
28 111113 7697 627 15 117923 10747 11
8 111776 7590 653 25 117960 10685 25
48 112461 714 2092 21 118002 10593 44
43 114579 10594 43 37 118065 8827 425
31 114660 6233 957 29 118519 8493 499
26 115627 10752 10 14 119056 10299 109
50 115639 9851 199 47 119192 10581 47
36 115855 10137 139 17 119246 9246 338
1 115997 9594 254 33 119614 4961 1287
27 116273 10551 53 38 120936 10059 164
5 116353 10646 33 7 121142 10754 10
49 116412 10610 40 45 121171 10658 31
19 116484 10660 29 16 121213 10738 14
18 116538 10622 38 35 121239 10259 120
4 116581 10329 100 32 121380 8628 483
10 116696 10596 43 11 121894 10010 176
24 116747 9648 245 3 122090 6697 924
20 116997 10773 6 40 123039 10790 2
12 117006 10794 1 34 123046 10747 12
39 117032 8565 477 6 123073 10519 63
13 117534 10020 167 30 123151 8319 563
9 117717 10613 40 22 123766 5722 1166
presented in Section 7.1. For = 38 , find out (), zk (1 k 4), and k

(0 k 4).
7.2 Discuss the pros and cons of energy-harvesting and wireless energy
transfer for a sensor node.
7.3 Lemma 7.1 shows the equivalence (in terms of their optimal objective
value) of a renewable energy cycle with a fully re-charged battery and a renew-
able energy cycle with partially re-charged battery. Whats the purpose of giv-
ing this equivalence in Lemma 7.1?
7.4 Give the second part of the proof so as to complete the proof of
Lemma 7.1.
181 7.11 Problems
Table 7.4 The case of clockwise traveling direction: Node visited along the path,
arrival time at each node, starting energy of each node in a renewable cycle, and
charging time at each node for the 50-node network.

the Path the Path
22 110676 5032 1166 9 117852 10613 40

30 111894 8032 563 13 117907 10023 167
6 112472 10489 63 39 118099 8588 477
34 112550 10741 12 12 118601 10795 1
40 112567 10789 2 20 118605 10773 6
3 112594 6301 924 24 118617 9668 245
11 113538 9944 176 10 118868 10600 43
32 113744 8461 483 4 118928 10339 100
35 114250 10222 120 18 119032 10626 38
16 114382 10734 14 19 119095 10664 29
45 114406 10648 31 49 119156 10615 40
7 114456 10751 10 5 119223 10650 33
38 114508 10012 164 27 119283 10558 53
33 114706 4676 1287 1 119357 9632 254
17 116024 9197 338 36 119613 10160 139
47 116368 10575 47 50 119770 9888 199
14 116443 10286 109 26 119972 10754 10
29 116590 8449 499 31 119992 6464 957
37 117118 8809 425 43 120986 10606 43
21 117561 10593 44 48 121056 1526 2092
25 117624 10685 25 8 123180 7927 653
15 117674 10747 11 28 123868 8059 627
23 117703 10793 2 46 124526 9472 305
44 117718 10605 42 41 124846 10637 37
2 117789 10611 41 42 124896 10754 11
7.5 How does Theorem 7.1 help us simplify the problem described in Sec-
tion 7.4? Assuming Theorem 7.1 is not provided, is there any change in prob-
lem OPT?
7.6 Give a formal proof of Corollary 7.1.
7.7 In Corollary 7.2, we claim that two opposite directions of the shortest
Hamiltonian cycle can achieve the same optimal objective. Explain why. In
the two optimal solutions corresponding to the two opposite directions, which
variables are changed and which are not? Justify your answer with the numer-
ical results in Section 7.9.
7.8 Property 7.1 shows that for an optimal solution, there exists at least one
bottleneck node in the network. Under a specific network, the WCV has two
opposite outgoing directions starting from its home service station (both follow
Figure 7.9 The energy behavior ei

of a sensor node (the 32th) in (J) Initial transient cycle First renewable cycle Second renewable cycle
the 50-node network during the Emax=10800
initial transient cycle and the

Ei=8628
first two renewable cycles.
Emin=540 t(s)
0
0 10752 11235 110628 121380 121863 221256 232008 232491 331884
(a) Traveling path is counter clockwise.

ei
(J) Initial transient cycle First renewable cycle Second renewable cycle
Emax=10800
Ei=8461
Emin=540 t(s)
0
0 3116 3599 110628 113744 114227 221256 224372 224855 331884
(b) Traveling path is clockwise.
Figure 7.10 The energy ei

(J) Initial transient cycle First renewable cycle Second renewable cycle
behavior of the bottleneck node
(48th node) in the 50-node Emax=10800
network. Traveling direction is

counter clockwise.
Ei=714
Emin=540 t(s)
0
0 1833 3925 110628 112461 114553 221256 223089 225181 331884
the shortest Hamiltonian cycle). In the two optimal solutions corresponding to

the two opposite directions, do we have the same bottleneck node(s)? Justify
your answer.
7.9 We consider a wireless sensor network consisting of ten nodes (see
Fig. 7.14). The sensor nodes are deployed over a square area of 100 m
100 m (see Table 7.7 for location of each sensor). The home service station
for the WCV is assumed to be located at (50, 70) (in m). The traveling speed
of the WCV is V = 5 m/s.
183 7.11 Problems
Figure 7.11 Comparison of data Y(m)

routing by our solution and that Sensor node
1000
by minimum energy routing for
the 50-node network.
500
Base station
X(m)
0
0 500 1000
(a) Data routing in our solution.
Y(m)
1000 Sensor node
500
Base station
X(m)
0
0 500 1000
(b) Minimal energy routing.
Find out the shortest Hamiltonian cycle for this network instance. For the
optimal cycle, what is the total traveling distance DTSP and the total traveling
time TSP ?
7.10 What is the benefit of reformulating problem OPT to problem OPT-R?
Table 7.5 Location and data rate Ri for each node in a 100-node network.
Node index Location (m) Ri (kb/s) Node index Location (m) Ri (kb/s)
1 (970, 383) 10 51 (295, 856) 7

2 (124, 85) 8 52 (306, 584) 10
3 (467, 734) 6 53 (106, 374) 8
4 (657, 332) 3 54 (594, 222) 7
5 (290, 840) 3 55 (283, 219) 10
6 (755, 372) 9 56 (155, 522) 1
7 (558, 828) 5 57 (1, 433) 3
8 (428, 177) 9 58 (284, 741) 10
9 (267, 130) 1 59 (551, 70) 8
10 (754, 880) 9 60 (871, 847) 5
11 (898, 44) 1 61 (42, 680) 7
12 (728, 687) 9 62 (905, 137) 5
13 (407, 734) 9 63 (131, 858) 4
14 (938, 437) 6 64 (834, 200) 3
15 (255, 380) 2 65 (800, 607) 4
16 (533, 980) 2 66 (918, 543) 1
17 (955, 399) 8 67 (137, 162) 5
18 (268, 440) 9 68 (505, 6) 4
19 (250, 157) 1 69 (405, 771) 10
20 (928, 326) 8 70 (174, 765) 6
21 (69, 314) 10 71 (575, 421) 9
22 (299, 895) 4 72 (606, 57) 5
23 (592, 247) 7 73 (214, 586) 5
24 (203, 311) 4 74 (520, 174) 9
25 (636, 409) 3 75 (989, 729) 10
26 (798, 708) 8 76 (490, 534) 6
27 (502, 144) 8 77 (695, 253) 10
28 (651, 871) 3 78 (411, 917) 8
29 (796, 83) 6 79 (35, 758) 6
30 (233, 462) 5 80 (293, 887) 10
31 (601, 30) 1 81 (801, 69) 9
32 (112, 753) 1 82 (347, 184) 10
33 (516, 700) 5 83 (83, 737) 7
34 (838, 215) 3 84 (511, 697) 4
35 (921, 680) 4 85 (367, 777) 5
36 (498, 557) 3 86 (739, 502) 10
37 (278, 851) 2 87 (525, 425) 1
38 (653, 559) 4 88 (805, 611) 2
39 (917, 902) 7 89 (817, 856) 5
40 (510, 420) 3 90 (189, 671) 6
41 (974, 358) 7 91 (124, 524) 1
42 (197, 489) 5 92 (821, 299) 7
43 (111, 256) 5 93 (638, 704) 9
44 (297, 929) 2 94 (16, 382) 2
45 (396, 467) 2 95 (896, 568) 5
46 (421, 254) 7 96 (515, 888) 3
47 (311, 431) 3 97 (545, 843) 10
48 (694, 703) 6 98 (606, 899) 7
49 (92, 402) 10 99 (760, 939) 3
50 (402, 182) 8 100 (855, 815) 2
185 7.11 Problems
Figure 7.12 An optimal Y(m)

traveling path for the WCV for
1000 Sensor node
the 100-node sensor network,
assuming traveling direction is
clockwise.
Mobile WCV
500
Base station
X(m)
0
0 Service station 500 1000
Table 7.6 The case of clockwise traveling direction: Node visited along the path,
arrival time at each node, starting energy of each node in a renewable cycle, and

the Path the Path
67 210727 10257 109 96 227822 10722 17

43 210855 9254 310 97 227850 9391 307
21 211179 10623 35 7 228162 7595 702
94 211232 10625 35 98 228880 10638 35
57 211278 10762 8 28 228926 9893 199
49 211305 10692 22 99 229151 10762 8
53 211333 9279 306 10 229171 10075 159
24 211661 8065 551 89 229344 10726 16
15 212230 8467 471 100 229371 9797 220
47 212716 4983 1181 60 229599 10613 41
18 213906 10698 21 39 229654 10687 25
30 213935 9012 364 75 229716 10581 48
42 214308 9784 207 35 229781 9415 305
56 214526 10787 3 95 230109 10561 53
91 214534 10778 4 66 230168 10790 2
61 214574 10682 24 14 230192 10720 18
79 214614 10681 24 17 230218 10723 17
83 214649 10274 107 1 230240 10287 113
32 214762 10365 89 41 230358 10716 19
Table 7.6 (Cont.)

the Path the Path
63 214873 10640 33 20 230388 8701 464

70 214926 9839 196 92 230874 7791 668
90 215141 9489 268 34 231559 10732 15
73 215427 8932 383 64 231577 9786 225
52 215828 9431 281 62 231822 10614 41
45 216139 1745 1876 11 231882 10772 6
76 218038 10737 13 81 231908 10715 19
36 218056 7075 775 29 231930 9728 239
38 218862 10152 135 72 232207 10740 13
86 219018 4116 1402 31 232226 10786 3
65 220444 8386 508 68 232249 10724 17
88 220952 9402 294 59 232281 10261 120
26 221266 9845 201 27 232419 10221 129
12 221482 10546 54 74 232556 8179 587
48 221543 10725 16 54 233160 10733 15
93 221570 9030 374 23 233180 7167 817
33 221969 1095 2072 77 234018 10563 53
84 224042 1197 2072 6 234098 6861 890
3 226126 8828 427 4 235009 6073 1075
13 226565 9705 237 25 236100 5792 1146
69 226809 10424 82 71 237259 6167 1066
85 226899 9884 199 87 238335 4268 1516
58 227115 10534 58 40 239855 10743 13
5 227193 9991 176 46 239906 10677 28
37 227372 10781 4 8 239950 10321 111
51 227380 10397 88 50 240066 10481 74
80 227473 10691 24 82 240152 10643 37
22 227499 10752 11 55 240203 9350 338
44 227517 10766 7 19 240555 10785 3
78 227547 9805 217 9 240564 10775 6
16 227791 10742 13 2 240600 10658 33
7.11 Why do we put data routing as a part of the optimization problem? Why
not simply use the minimum energy routing?
7.12 What is the purpose of using piecewise linear approximation for the
quadratic terms i2 ? How do we ensure that such an approximation can guar-
antee (1 )-optimality?
7.13 After solving problem OPT-L, why is it necessary to construct a feasible
solution to problem OPT-R (and problem OPT)? Verify the feasibility of the
constructed solution (in Section 7.7.3).
187 7.11 Problems
Figure 7.13 Piecewise linear

Y
approximation for a continuous 1
function () = 2 .
()
()
0 1 1 3 1 X
4 2 4
Figure 7.14 A ten-node sensor Y(m)

network.
100
Service station
Sensor node
50
X(m)
0
0 50 100
7.14 We consider a small WSN consisting of three nodes (see Fig. 7.15). The
length of the shortest Hamiltonian cycle is 120 m for this network instance. The
traveling speed of the WCV is 1.2 m/s. Each sensor node has a battery capacity
of 10 KJ and is fully charged initially. The minimum energy at a sensor node
battery for it to be operational is 500 J. The wireless energy transfer rate is 5
W. We are looking for a solution within 1% from optimum.
Answer the following questions:
(a) How can we set m, i.e., the number of segments that we use in the piecewise
linear approximation such that 99% optimality is achieved?
(b) Use piecewise linear approximation and derive a relaxed linear formulation
in the forms of problem OPT-L.
Table 7.7 Sensor node location in a ten-node network.
Node Index Location (m) Node Index Location (m)
1 (60, 30) 6 (35,40)

2 (70, 65) 7 (45,55)
3 (90, 80) 8 (80,60)
4 (60, 40) 9 (20,60)
5 (45, 35) 10 (40,25)
Figure 7.15 A three-node Y(m)

sensor network. 50
Sensor node
Service station X(m)

0
0 50
Table 7.8 Location and data rate Ri for each node in a 20-
node network.
Node Location Ri Node Location Ri

index (m) (kb/s) index (m) (kb/s)
1 (600, 500) 8 11 (800, 700) 3

2 (700, 550) 9 12 (470, 700) 5
3 (560, 560) 1 13 (650, 250) 8
4 (400, 900) 9 14 (300, 100) 9
5 (560, 300) 6 15 (520, 540) 7
6 (100, 200) 4 16 (750, 150) 10
7 (590, 480) 3 17 (220, 200) 7
8 (800, 480) 6 18 (850, 300) 4
9 (200, 600) 5 19 (980, 150) 6
10 (400, 250) 10 20 (50, 950) 2
7.15 Why is it necessary to construct the initial transient cycle?

7.16 We consider a randomly generated WSN consisting of 20 nodes.
Table 7.8 gives the location of each node and its data rate for a 20-node net-
work. Refer to Section 7.9.1 for parameter setting.
The shortest Hamiltonian cycle is shown in Fig. 7.16. For this optimal cycle,
DTSP = 4046 m and TSP = DTSP /V = 809.2 s. For the target = 0.01, by
Theorem 7.2, we have m = 4. In our solution, we have cycle time = 31.47
189 7.11 Problems
Figure 7.16 An optimal Y(m)

traveling path for the 20-node
1000
sensor network. Only clockwise Mobile WCV
traveling direction is shown.
Sensor node
500
Base station
X(m)
0
0 Service station 500 1000
Table 7.9 The case of clockwise traveling direction: Node visited along the path, arrival
time at each node, renewable cycle starting energy of each node, and charging time
at each node for the 20-node network.
Node Visited ai (s) Ei (J) i (s) Node Visited ai (s) Ei (J) i (s)
Along the Path Along the Path
6 113337 10655 29 2 117427 9331 306

17 113390 10280 104 11 117768 10336 97
9 113574 5769 1018 8 117909 9400 293
20 114668 6280 923 18 118239 5765 1063
4 115661 7550 668 19 119342 9473 281
12 116372 8299 517 16 119669 8034 589
15 116922 10758 9 13 120287 10316 103
3 116940 10402 82 5 120411 1197 2091
7 117039 9608 247 10 122535 7361 754
1 117291 10252 114 14 123326 7823 657
hours, the vacation time vac = 28.48 hours, and the objective vac = 90.50%.
Table 7.9 shows nodes visited along the path of clockwise traveling direction,
arrival time at each node, renewable cycle starting energy of each node, and
Assuming the WCV follows the clockwise traveling direction, answer the
following questions:
(a) Based on Table 7.9, how to determine the energy consumption rate pi at
sensor node i? In particular, what are p5 , p6 , and p14 ?
(b) In the initial transient cycle, the WCV re-charges sensor node i at the charg-
ing rate of ui . How to determine the value of ui for sensor node i? In particular,
what are u5 , u6 , and u14 ?
(c) Theorem 7.3 shows that the constructed initial cycle is a feasible transient
cycle. Demonstrate this point at the 18th node. Draw the energy behavior of
this specific node during the initial transient cycle and the first two renewable
cycles.
(d) Recall that a bottleneck node in the network is defined as a node with
its energy dropping to Emin during a renewable energy cycle. Which node is
the bottleneck node in this network? Justify your answer and draw the energy
behavior of the bottleneck node during the initial transient cycle plus the first
two renewable cycles.
CHAPTER
8 Approximation algorithm and its

applications - Part 1
Do not follow where the path may lead. Go, instead, where there is no path and
leave a trail.
Ralph Waldo Emerson
8.1 Review of approximation algorithms
Recall that Part II of this book focuses on methods for near-optimal and
approximation solutions. Specifically, Chapters 5 to 7 follow the OR optimiza-
tion approach to develop a (1 )-optimal solution. In this and the following
chapters, we will show how to develop the (1 )-optimal solution by follow-
ing the CS algorithm design approach.
We first give a brief overview of the so-called approximation algorithms.
Such algorithms are designed to offer solutions that can approximate the
unknown optimal solution according to certain benchmark performance crite-
ria. In particular, in wireless network research, the most popular approximation
algorithms can be classified as constant-factor approximation algorithms and
(1 )-optimal approximation algorithms.
Constant-factor approximation algorithms. We use a maximization

problem with a positive optimal objective value OPT (unknown) as an exam-
ple. If we can prove that the obtained feasible solution achieves an objective
value that is at least c OPT, where c < 1 is a constant, then the designed
algorithm is a constant-factor approximation algorithm. Likewise, for a
minimization problem, if we can prove that the obtained feasible solution
achieves an objective value that is at most c OPT, where c > 1 is a constant,
then the designed algorithm is a constant-factor approximation algorithm.
191
192 Approximation algorithm and its applications - Part 1
(1)-optimal (or (1 + )-optimal) approximation algorithms. If the

approximation factor can be set arbitrarily close to 1, then the approxi-
mation algorithm is a (1 )-optimal (or (1 + )-optimal) approximation
algorithm. That is, for a maximization problem and any small constant ,
0 < 1, if the algorithm can find a solution with an objective value that is
at least (1 )OPT, then this algorithm is (1 )-optimal. For a minimiza-
tion problem and any small constant , if the algorithm can find a solution
with an objective value that is at most (1 + )OPT, then this algorithm is
(1 + )-optimal.
Both of the above two types of approximation algorithms are common in the
literature to solve wireless network problems. The constant-factor approxima-
tion algorithms would only be useful if c is close to 1. But unfortunately, many
results in the literature offer results that may be far away from 1. Although
such results may still be of theoretical value, they are hardly useful in practice
(in the sense of offering an accurate performance benchmark for the design of
distributed networking algorithms and protocols).
Our preference is toward the design of (1 )-optimal approximation
algorithms, which will be presented in a case study in this chapter. As
expected, such algorithms are intellectually challenging and require much
novelty in their design. Nevertheless, should we be able to design such an
algorithm, then both its theoretical significance and practical value would be
assured.
The case study in this chapter is a classic and fundamental problem on base
station placement in a wireless sensor network (WSN). The setting of this prob-
lem is similar to that in Chapter 2. The problem here is to find the optimal
location for the base station so that the network lifetime (until any sensor node
runs out of energy) is maximized.
8.2 Case study: The base station placement problem
As discussed in Chapter 2, an important performance metric for a WSN is

the so-called network lifetime. Given that energy expenditure at a node for
transmitting data to another node directly depends on the physical distance
between these two nodes, network lifetime is therefore highly dependent upon
the physical topology of the network.
In this case study, we focus on base station location, which, as expected, has
a significant impact on network lifetime performance. Specifically, we consider
the following problem. Suppose each node i producing sensing data at a rate
of ri in a WSN, where all the sensing data are to be forwarded to the base
station (via multi-hop if necessary): where should we place the base station in
this WSN so that the network lifetime is maximized?
193 8.2 Case study: The base station placement problem
We will show that this problem can be formulated into a nonlinear program-
ming (NLP), which is nonconvex and NP-hard in general [46]. One approach
is to employ the branch-and-bound framework in Chapter 5. But that solu-
tion approach is tedious and not elegant. Instead, we will design an elegant
approximation algorithm that can guarantee (1 )-optimal network lifetime
performance.
The (1 )-optimal approximation algorithm that we are going to present
is based on several techniques, which makes it possible to reduce an infinite
search space to a finite-element search space for the base station location. The
main idea of the approximation algorithm is to exploit a clever way of dis-
cretizing the cost parameter associated with energy consumption into a geo-
metric sequence with tight upper and lower bounds. As a result, we can divide
a continuous search space into a finite number of subareas. By further exploit-
ing the cost property of each subarea, we can represent each subarea with a
so-called fictitious cost point (FCP), which is an N -tuple cost vector with
each element representing an upper bound on the cost for a corresponding sen-
sor node in the network. Based on these ideas, we can successfully reduce
an infinite search space for the base station location into a finite number of
points for each of which we can solve an LP problem to find an achiev-
able network lifetime and data routing solution. By comparing the achievable
network lifetime obtained among all the FCPs, we show that the largest is
(1 )-optimal. We also show that placing the base station at any physical
point in the subarea corresponding to the best FCP is (1 )-optimal. We
analyze the complexity of the approximation algorithm and show that it is
polynomial.
The remainder of this chapter is organized as follows. Section 8.3 describes
the network model used in the case study and formally states the base station
placement problem. In Section 8.4, we show that for a given base station loca-
tion, the maximum achievable network lifetime and the corresponding optimal
routing can be found via a single LP problem. To search the best base sta-
tion location, in Section 8.5, we narrow down the search space to the small-
est enclosing disk (SED) that covers all the sensor nodes in the network. In
Section 8.6, we divide the continuous search space of the SED into a finite
number of subareas and represent each subarea by an FCP. Thus, we can find
the best FCP corresponding to the maximum network lifetime. We then show
that by placing the base station at any point in the subarea corresponding to
the best FCP, we obtain a (1 )-optimal network lifetime. In Section 8.7, we
summarize all the steps as an algorithm. We also give an example for illus-
tration. In Section 8.8, we prove the correctness of the algorithm and ana-
lyze its complexity. In Section 8.9, we give some additional numerical exam-
ples illustrating the efficacy of the algorithm. Section 8.10 summarizes this
chapter.
8.3 Network model and problem description

8.3.1 Network model
We consider a static sensor network consisting of a set N of sensor nodes
deployed over a two-dimensional area. The location of each sensor node is
fixed and the initial energy at sensor node i is denoted as ei . Each sensor node
i generates data at a rate ri . We assume that there is a single base station to be
deployed in the area to collect sensing data.
In this chapter, we focus on the energy consumption due to communications
(i.e., data transmission and reception). Suppose that sensor node i transmits
data to sensor node j with a rate of fij (b/s). Then we model the transmission
power at sensor node i as
utij = Cij fij . (8.1)
Here, Cij is the cost associated with link i j and can be modeled as
Cij = 1 + 2 dij , (8.2)
where 1 and 2 are constant coefficients, dij is the physical distance between
sensor nodes i and j , and is the path-loss index, with 2 4.
The power consumption at the receiving sensor node i can be modeled as
k=i

uri = fki , (8.3)
kN
where fki (b/s) is the incoming bit-rate received by sensor i from sensor k and
is a constant coefficient.
In this chapter, we assume that the interference from different transmissions
has been effectively avoided by appropriate MAC layer scheduling. For low bit
rate and deterministic traffic pattern considered in this chapter, a contention-
free MAC protocol is not hard to design and we omit its discussion in this
chapter. Table 8.1 lists the notation used in this chapter.
8.3.2 Problem description

We aim to investigate how to optimally place a base station to collect data in
a WSN so that the network lifetime can be maximized. Network lifetime is
defined as the time until any sensor node first uses up its energy. To achieve
optimality, we allow flow splitting, just as we did in previous chapters. Also,
power control at a node is allowed, as modeled in (8.1) and (8.2).
Assume that base station B is located at a point p. Denote (xB , yB ) as the
position of point p and let T be the network lifetime. A feasible flow routing
solution that achieves this network lifetime T should satisfy both flow balance
and energy constraints at each sensor node. These constraints can be formally
stated as follows. Denote fij and fiB as the data rates from sensor node i to
195 8.3 Network model and problem description
Table 8.1 Notation.
Symbol Definition
A The search space for the base station, which can be the smallest
enclosing disk to cover all sensor nodes
Am The mth subarea in the search space
B The base station
Cij Power consumption coefficient for transmitting data from sensor i
to sensor j
ciB (p) Power consumption coefficient for transmitting data from sensor i
to base station B at point p
CiBmin , C max Lower and upper bounds of c (p)
iB iB
C[h] The transmission cost for the hth circle
dij (or diB ) Distance between sensor i and sensor j (or base station B)
ei Initial energy at sensor i
fij (or fiB ) Data rate from sensor i to sensor j (or base station B)
Hi Total number of circles for discretization at sensor node i for a given
K Total number of circles for discretization for a given
M Total number of subareas for discretization for a given
N Set of sensor nodes in the network
N = |N |, number of sensor nodes in the network
OA The center of the smallest enclosing disk A
pm Fictitious cost point (FCP) representation for the mth subarea
popt The best base station location
p The best location among M FCPs
p A point in the subarea corresponding to p
ri Sensing data rate produced at sensor i
RA The radius of the smallest enclosing disk A
Tm Maximum achievable network lifetime by placing the base station at pm
Topt Optimal network lifetime achieved by placing the base station at popt
T = max{Tm : m = 1, 2, . . . , M}
T (1 )-optimal network lifetime achieved by p
Vij (or ViB ) Total data volume from sensor i to sensor j (or base station B)
(xi , yi ) Location of sensor node i
Path-loss index
1 , 2 Constant coefficients in transmission power modeling
Desired approximation error, 0 < 1
Power consumption coefficient for receiving data
opt The best flow routing solution when the base station is at popt
Flow routing solution when the base station is at p
Flow routing solution when the base station is at p
sensor node j and base station B, respectively. Then the flow balance for each
sensor node i is
k=i
j =i

fki + ri = fij + fiB ,
kN j N
i.e., the sum of total incoming flow rates plus the self-generated data rate is
equal to the sum of the total outgoing flow rates. The energy constraint for
each sensor node i is
k=i
j =i

fki T + Cij fij T + ciB (p) fiB T ei ,
kN j N
i.e., total consumed energy due to receiving and transmission over time T can-
not exceed its initial energy ei . By (8.2), we have
0
ciB (p) = 1 + 2 (xB xi ) + (yB yi )
2 2 ,
which is a nonlinear function of base station location (xB , yB ).

Our objective is to maximize the network lifetime T under the flow balance
and energy constraints, i.e.,
Maximize T
j =i
k=i

subject to fij + fiB fki = ri (i N ) (8.4)
j N kN
k=i
j =i

fki T + Cij fij T + ciB (p)fiB T ei (i N ) (8.5)
kN j N
0
ciB (p) 2 (xB xi )2 + (yB yi )2 = 1 (i N )
(xB , yB ) A, T , fij , fiB 0 (i, j N , i = j ),

where A is an area containing all possible base station locations. This opti-
mization problem is in the form of a nonconvex programming, which is NP-
hard in general [46].
8.4 Optimal flow routing for a given base station location
As discussed earlier, the maximum network lifetime depends on both the base
station location and data routing. To begin with, we show that for a given base
station location, we can find the maximum network lifetime and optimal rout-
ing via a single LP problem. Here, the objective function is the network lifetime
T and the constraints are given by (8.4) and (8.5). Multiplying both sides of
(8.4) by T and denoting
Vij = fij T and ViB = fiB T , (8.6)
where Vij (or ViB ) can be interpreted as the total data volume from sensor
node i to sensor node j (or base station B) over time T , we have the following
formulation:
197 8.5 Search space for base station location
Maximize T
k=i
j =i

subject to Vki + ri T Vij ViB = 0 (i N )
kN j N
k=i
j =i

Vki + Cij Vij + ciB (p)ViB ei (i N )
kN j N
T , Vij , ViB 0 (i, j N , i = j ).
Note that for a given base station location, the ciB (p)-values are constants.
Therefore, the above formulation is in the form of an LP problem. Once we
solve the above LP problem, we can obtain an optimal routing solution for fij
V
and fiB by fij = Tij and fiB = VTiB .
8.5 Search space for base station location
Although for a given base station location we can find the corresponding max-
imum network lifetime via a single LP problem, it is not possible to examine
all (infinite) points in the two-dimensional plane and select the point having
the maximum network lifetime.
As a first step, we show that it is only necessary to consider points inside the
so-called smallest enclosing disk (SED) [167],1 which is a unique disk with
the smallest radius that contains all the N sensor nodes in the network and can
be found in O(N) time [108]. This is formally stated in the following lemma:
Lemma 8.1
To maximize network lifetime, the base station location must be within the
smallest enclosing disk A that covers all the N sensor nodes in the network.
Proof. The proof is based on contradiction. That is, if the base station location
is not in the SED, then its corresponding network lifetime cannot be maximum.
To see this, assume that the optimal base station location is at point p, which
is outside the SED A (see Fig. 8.1). Denote OA as the center of the SED. Let
q be the intersecting point between the line segment [p, OA ] and the circle
1 In fact, we can consider points in an even smaller area, i.e., the convex hull of all sensor nodes.
However, using the convex hull does not reduce the order of complexity of our algorithm. On
the other hand, the use of the SED simplifies the discussion.
Figure 8.1 A schematic diagram

showing that an optimal base
station location must be within
the SED.
i
p
q
OA
RA
of the SED. Then for any sensor node i (all in A), we have diq < dip . Con-
sequently, Ciq < Cip . As a result, we can save transmission energy for every
sensor node i N by relocating p to q. This saving in energy at each node
increases network lifetime, which shows that point p cannot be the optimal
location to maximize network lifetime.
We have thus narrowed down the search space for base station B from a
two-dimensional plane to the SED A. However, the number of points in A
remains infinite. It is tempting to partition A into small subareas (e.g., a grid-
like structure), A1 , A2 , . . ., up to say AM , i.e.,

M
A= Am .
m=1
When each subarea is sufficiently small (i.e., M is sufficiently large), we can

use some point qm Am to represent Am , m = 1, 2, . . . , M. By solving an
LP problem on each of the M points, we can select the best location among
all points and obtain a good solution for the base station placement. However,
such an approach is a heuristic at best, and does not provide any theoretical
guarantee on performance.
The key to providing a theoretical guarantee on performance is to partition
each subarea in such a way that tight bounds on transmission cost can be guar-
anteed for any point in the subarea. If this is possible, then we may be able
to exploit such a property and design an approximation algorithm that yields
provably (1 )-optimal network lifetime performance. The goal of this chap-
ter is to develop such an algorithm having a guaranteed performance rather
than merely a good heuristic. In the following section, we develop a novel
technique to partition the SED A into subareas where each subarea can be
represented by a special point with a set of tight bounds. Then we show how
to design a (1 )-optimal approximation algorithm based on these special
points.
199 8.6 Subarea division and fictitious cost points
Figure 8.2 A sequence of

circles with increasing costs with 1
center at node 4.
OA RA
4
D[1] D[2] D[3]

3
8.6 Subarea division and fictitious cost points

8.6.1 Subarea division
The subarea division (with guaranteed performance bounds) hinges upon a
clever discretization of a cost parameter. A close look at the energy constraint
in (8.5) suggests that the location of the base station is embedded in the cost
parameters ciB (p). In other words, if we can discretize these cost parameters,
we may also discretize the location for the base station.
Since the search space is narrowed down to the SED A, we can limit the
range for the distance between a sensor node i to the possible location for the
base station. Denote OA and RA as the origin and radius of the SED A. For
each sensor node i N , denote Di,OA as the distance from sensor node i to
the origin of disk A (see node 4 in Fig. 8.2 as an example). Denote by DiB min
max
and DiB the minimum and maximum distances between sensor node i and
any possible location for the base station B, respectively. Then we have
min
DiB = 0,
max
DiB = Di,OA + RA .
min max min max

Corresponding to DiB and DiB , denote by CiB and CiB the minimum and
maximum cost values between sensor node i and base station B, respectively.
Then by (8.2), we have
min
CiB = 1 , (8.7)
max
CiB = 1 + 2 (DiB ) = 1 + 2 (Di,OA + RA ) .
max
(8.8)
Given the range of diB [DiBmin max

, DiB ] = [0, Di,OA + RA ] for each sensor
node i, we now show how to partition the disk A into a finite number of sub-
areas with the distance of each subarea to sensor node i meeting some tight
bounds. Specifically, from a sensor node i, we draw a sequence of circles cen-
tered at this sensor node, each with increasing radius D[1], D[2], . . . , D[Hi ]
Figure 8.3 An example of

subareas within disk A that are
obtained by intersecting arcs
from different circles.
1
q3
q1
q2
4
corresponding to costs C[1], C[2], . . . , C[Hi ], which are defined with the fol-
lowing geometric property:
C[h] = CiB
min
(1 + )h = 1 (1 + )h (1 h Hi ). (8.9)
This geometric sequence C[h] (with a factor of (1 + )) is carefully chosen

and will offer tight performance bounds for any point in a subarea (more on
this later). The number of required circles Hi can be determined by having the
last circle in the sequence (with radius D[Hi ]) to completely contain disk A,
i.e., D[Hi ] DiBmax
, or equivalently,
C[Hi ] CiB
max
.
That is, we can determine Hi as follows:

/ max min 1 ln 1 + 2
(D i,O + RA )
ln(CiB /CiB ) 1 A
Hi = =
.
(8.10)
ln(1 + ) ln(1 + )
For example, for node 4 in Fig. 8.2, we have H4 = 3, i.e., D[3] is the circle
centered at node 4 that will completely contain the disk. As a result, with sensor
node i as the center, we have a total of Hi circles, each with cost C[h], h =
1, 2, . . . , Hi .
The above partitioning of the SED A is with respect to a specific node i. We
now perform the above process for all sensor nodes. These intersecting circles
will cut disk A into a finite number of irregular subareas, with the boundaries
of each subarea being either an arc (with a center at some sensor node i and
some cost C[h], 1 h < Hi ) or an arc segment of the SED A. As an example,
the SED A in Fig. 8.3 is now cut into 28 irregular subareas.
We now claim that under this subarea partitioning technique, for any point
in a given subarea, its cost to each sensor node in the network can be tightly
bounded quantitatively. This is because, with respect to each sensor node i,
201 8.6 Subarea division and fictitious cost points
a subarea Am must be enclosed within some arc centered at sensor node i.

Denote the index of this arc (w.r.t. sensor node i) as hi (Am ). So when the base
station B is at any point p Am , we have
C[hi (Am ) 1] ciB (p) C[hi (Am )], (8.11)
C[hi (Am )]
where we define C[0] = CiB min
= 1 . Since C[h i (Am )1]
= 1 + by (8.9), we
have very tight lower and upper bounds for ciB (p). The reader may now have
a better appreciation of the benefit of the discretization technique for costs and
distances.
8.6.2 Fictitious cost point

We now introduce a novel concept called fictitious cost point (FCP). It will be
used to represent an upper bound on the cost for any point in a subarea Am ,
m = 1, 2, . . . , M.
Definition 8.1
Denote the FCP for subarea Am (m = 1, 2, . . . , M) as pm , which is repre-
sented by an N-tuple vector with its ith element (i = 1, 2, . . . , N) being an
upper bound on the cost for any point in subarea Am corresponding to the
ith sensor node in the network.
That is, the N -tuple cost vector for FCP pm is [c1B (pm ), c2B (pm ), . . .,
cNB (pm )], with the ith element ciB (pm ) being
ciB (pm ) = C[hi (Am )], (8.12)
where hi (Am ) is determined by (8.11).
As an example, the FCP for subarea with corner points {q1 , q2 , q3 } in
Fig. 8.3 can be represented by a 4-tuple cost vector [c1B (pm ), c2B (pm ),
c3B (pm ), c4B (pm )] = [C[2], C[3], C[2], C[3]], where the first component
C[2] represents an upper bound on the cost for any point in this subarea to
sensor node 1, the second component C[3] represents an upper bound on the
cost (which is loose here) for any point in this subarea to sensor node 2, and
so forth.
We emphasize that the reason we use the word fictitious is that an FCP pm
cannot be mapped to a physical point within the corresponding subarea Am .
This is because there does not exist a physical point in subarea Am that has its
costs to all the N sensor nodes equal (one by one) to the respective N -tuple cost
vector embodied by pm simultaneously. As an example, any point within the
dark subarea bounded by corner points q1 , q2 , and q3 cannot have its costs cor-
respond to the four sensor nodes in the network equal to the respective element
in [C[2], C[3], C[2], C[3]] simultaneously, where [C[2], C[3], C[2], C[3]] is
the cost vector for the FCP that represents this subarea.
The following important property for FCP pm will be used in the proof of
the (1 )-optimality guarantee of the approximation algorithm.
Property 8.1
For any point p Am and the corresponding FCP pm , we have
ciB (pm ) (1 + )ciB (p).
Proof. By the definitions of FCP pm (see (8.12)) and C[h] (see (8.9)), we have
ciB (pm ) = C[hi (Am )] = (1 + ) C[hi (Am ) 1] (1 + ) ciB (p),
where the inequality follows from (8.11).
8.7 Summary of algorithm and example
By discretizing the cost parameters and the corresponding distances, we have

partitioned the search space (SED A) into a finite number of M subareas. By
introducing the concept of FCPs, we can represent each subarea with a single
point. As a result, we can now readily apply the LP formulation of Section 8.4
to examine each FCP and thereby select the FCP that offers the maximum
network lifetime. The complete approximation algorithm is outlined in Algo-
rithm 8.1. The correctness proof of its (1 )-optimality guarantee is given in
Section 8.8.
Example 8.1
We use a small three-node network to illustrate the steps of the approxima-
tion algorithm. The location, data rate, and initial energy for each sensor
are specified in Table 8.2, where the units of distance, rate, and energy are
all normalized. Also, we set = 2, 1 = 1, 2 = 0.5, and = 1 under the
normalized units. For illustration, we set the error bound to = 0.2.2
(1) We identify SED A with origin OA = (0.61, 0.57) and radius RA =

0.51 (see Fig. 8.4).
(2) We first have Di,OA = RA = 0.51 for each node i, 1 i 3. We then
find the lower and upper bounds on ciB for each node i, 1 i 3, as
follows:
min
CiB = 1 = 1,
max
CiB = 1 + 2 (Di,OA + RA ) = 1 + 0.5 (0.51 + 0.51)2 = 1.52.
203 8.7 Summary of algorithm and example
Algorithm 8.1 A (1 )-approximation algorithm

1. Find the smallest enclosing disk A that covers all the N nodes.
2. Within A, compute the lower and upper cost bounds CiB min max
and CiB for each
node i N by (8.7) and (8.8).
3. For a given > 0, define a sequence of costs C[1], C[2], . . . , C[Hi ] by
(8.9), where Hi is calculated by (8.10).
4. At each node i, draw a sequence of (Hi 1) circles centered at node i with
increasing radius corresponding to cost C[h], h = 1, 2, . . . , Hi 1. The
intersection of these circles within disk A will partition A into M subareas
A 1 , A2 , . . . , AM .
5. For each subarea Am , 1 m M, define an FCP pm by an N -tuple cost
vector [c1B (pm ), c2B (pm ), . . ., cNB (pm )], where ciB (pm ) is defined in (8.12).
6. For each FCP pm , 1 m M, apply LP in Section 8.4 to obtain the achiev-
able network lifetime Tm .
7. Select the FCP p that offers the maximum network lifetime among these
M FCPs. The base station can be placed at any point p within the subarea
corresponding to p .
8. For the chosen point p , apply LP in Section 8.4 and obtain a (1 )-
optimal network lifetime T .
(3) For each node i, 1 i 3, we find

ln 1 + 21 (Di,OA + RA ) ln 1 + 0.5
1 (0.51 + 0.51)
2
Hi =

=

= 3,

ln(1 + ) ln(1 + 0.2)
and
C[1] = 1 (1 + ) = 1 (1 + 0.2) = 1.20,

C[2] = 1 (1 + )2 = 1 (1 + 0.2)2 = 1.44,
C[3] = 1 (1 + )3 = 1 (1 + 0.2)3 = 1.73.
(4) We draw circles centered at each node i, 1 i 3, with cost C[h],

1 h < Hi = 3, to partition the original disk A into 16 subareas
A1 , A2 , . . . , A16 .
(5) We define an FCP pm for each subarea Am , 1 m 16. For
example, for FCP p1 , we define a 3-tuple cost vector as [c1B (p1 ),
c2B (p1 ), c3B (p1 )] = [C[1], C[3], C[2]] = [1.20, 1.73, 1.44].
(6) We apply the LP formulation of Section 8.4 to these 16 FCPs and obtain
the network lifetime for each FCP.
(7) Since the FCP p = p9 has the maximum achievable network lifetime
of 226.47 among all the 16 FCPs, we can place the base station at any
point in the subarea A9 , e.g., p = (0.6, 0.6).
Table 8.2 Sensor locations, data rate, and initial

energy for the example sensor network.
Node index Location Data rate Initial energy
1 (0.1, 0.5) 0.8 390

2 (1.1, 0.7) 1.0 400
3 (0.4, 0.1) 0.5 130
3 6
1 7 11
4 14
0.8 8
1 12
0.6
Y 15
9
0.4 2
5
13
0.2 16
10
0
0 0.2 0.4 0.6 0.8 1
X
Figure 8.4 The SED is partitioned into 16 subareas for the three-node example.
(8) We apply the LP formulation of Section 8.4 to p and obtain a (1 )-

optimal network lifetime T = 227.07. This completes the algorithm.
8.8 Correctness proof and complexity analysis
In this section, we give a formal proof that the solution obtained by Algo-
rithm 8.1 is (1 )-optimal and analyze its complexity.
Denote popt as the optimal location for base station placement, Topt and opt
as the corresponding maximum network lifetime and flow routing solution, all
of which are unknown.
Denote p as the best FCP among the M FCPs pm , m = 1, 2, . . . , M,
based on their network lifetime performance. Denote by T and the
2 This is used here to simplify the illustration of each step. In Section 8.9, we use = 0.05 for
all computations.
205 8.8 Correctness proof and complexity analysis
corresponding maximum network lifetime and flow routing solution, i.e.,

T = max{Tm : m = 1, 2, . . . , M}.
Based on Algorithm 8.1, Step 7, we choose a physical point p in the sub-
area corresponding to p for base station placement. For point p , denote the
network lifetime as T and the corresponding flow routing solution as .
Our roadmap for the proof is as follows. In Theorem 8.1, we prove that T
for the best FCP p is within (1 ) of the optimum, i.e., T (1 )Topt .
Then, in Theorem 8.2, we show that for the physical point p , its corre-
sponding network lifetime T is also within (1 ) of the optimum, i.e.,
T (1 )Topt .
Theorem 8.1
T (1 )Topt .
To prove Theorem 8.1, we first present the following lemma, which is a

general case of the theorem:
Lemma 8.2
For any given base station location p and corresponding optimal routing
solution and achievable network lifetime T (obtained via the stated LP
formulation), denote Am as the subarea that contains p for a given . Then
for the corresponding FCP pm , its achievable network lifetime Tm is at least
(1 ) of T , i.e., Tm (1 )T .
Proof. Instead of considering the optimal routing solution for FCP pm , we use
the same routing on pm , which is clearly suboptimal. That is, denoting Tm
as the network lifetime for FCP pm under , we have Tm Tm . Then we only
need to show that Tm (1 )T .
To show Tm (1 )T , we compute the total consumed energy on node
i N under for FCP pm and at time (1 )T , which is
k=i
j =i

fki (1 )T + Cij fij (1 )T + ciB (pm )fiB (1 )T
kN j N
k=i
j =i

< fki T + Cij fij T + (1 + )ciB (p)fiB (1 )T
kN j N
k=i
j =i

< fki T + Cij fij T + ciB (p)fiB T ei .
kN j N
The first inequality holds via Property 8.1. The last inequality holds by the
energy constraint corresponding to the routing solution for the point p. Thus,
the network lifetime Tm for FCP pm under the routing solution is at least
(1 )T . Thus, we have Tm Tm (1 )T .
With Lemma 8.2, we are ready to prove Theorem 8.1 as follows:
Proof. Consider the special case of Lemma 8.2 where the given base station
location p is the optimal location popt , with the corresponding optimal data
routing solution opt and network lifetime Topt . Following the same approach
as in Lemma 8.2, we can find a corresponding subarea Am that contains the
point popt with the corresponding FCP pm . As a result, for FCP pm , we have
Tm (1 )Topt . Thus, for the best FCP p among all the FCPs, we have
T Tm (1 )Topt .
Theorem 8.1 guarantees that the best network lifetime among the M FCPs
is at least (1 ) of Topt . Now, consider a point p in the subarea represented
by the best FCP p . We have the following theorem:
Theorem 8.2
T (1 )Topt .
Proof. Denote T as the network lifetime for point p under the same routing
solution for FCP p . Since is a suboptimal routing for p , we have T
T . Thus, to show T (1 )Topt , we only need to show that T T
(1 )Topt , where the second inequality follows from Theorem 8.1.
To establish T T , we compute the total consumed energy on node i N
under for the point p at time T , which is given by
k=i
j =i

fki T + Cij fij T + ciB (p )fiB T
kN j N
k=i
j =i

fki T + Cij fij T + ciB (p )fiB T ei .
kN j N
The first inequality holds by (8.11) and (8.12). The second inequality holds by
the energy constraint on p under the routing solution . Thus, the network
lifetime T for location p under is at least T . As a result, the maximum
network lifetime T for location p is at least T T (1 )Topt .
The complexity of Algorithm 8.1 can be measured by the number of LPs that
need to be solved, which is equal to the total number of subareas M. Hence,
we compute M.
The boundary segments of each subarea are defined by either an arc centered
at some sensor node i (with some cost C[h], 1 h < Hi , with Hi being given
by (8.10)), or an arc of the disk A. Since there are Hi 1 circles radiating from
each sensor node i and one circle for the disk A, the total number of circles
Table 8.3 Each nodes location, data generation rate, and initial energy for a ten-node
network.
Location Rate Initial energy Location Rate Initial energy
(0.81, 0.86) 0.7 390 (0.48, 0.22) 0.1 300

(0.25, 0.71) 0.4 400 (0.53, 0.16) 0.8 410
(0.47, 0.44) 1.0 440 (0.66, 0.52) 0.2 210
(0.28, 0.03) 0.6 330 (0.91, 0.86) 0.1 320
(0.25, 0.36) 0.2 440 (0.44, 0.21) 0.9 330

is K = 1 + iN (Hi 1). The maximum number of subareas M that can be
obtained by K circles is upper bounded by
M K 2 K + 2. (8.13)
The verification of (8.13) is left as a homework exercise. We hence have

# $2 #% & $
N 2
M = O(K ) = O
2
Hi = O .

iN
In this section, we apply the approximation algorithm to various network

topologies and use numerical results to demonstrate its efficacy. The units of
distance, rate, and energy are all normalized appropriately. The normalized
parameters in the energy consumption model are 1 = 2 = = 1, and we
set the path-loss index = 2.
We consider four randomly generated networks consisting of ten 20, 50, and
100 nodes deployed over a 1 1 square area. In all cases, the targeted accuracy
for the approximation algorithm is 0.95-optimal, i.e., = 0.05.
The network setting (location, data rate, and initial energy for each node) for
the ten-node network is specified in Table 8.3. By applying Algorithm 8.1, we
find that the FCP with the cost vector [1.05, 1.28, 1.05, 1.22, 1.16, 1.05, 1.05,
1.05, 1.41, 1.05] has the maximum network lifetime T = 357.49, which is at
least 95% of the optimal value. By placing the base station at a point in the
corresponding subarea, e.g., at point (0.59, 0.31), the network lifetime is T =
359.17 > T . This network lifetime is also at least 95% of the optimal value.
The flow routing solution is displayed in Fig. 8.5, where a circle represents a
sensor node and a star represents the location of the base station (0.59, 0.31).
The network setting for a 20-node network (with location, data rate, and
initial energy for each of the 20 sensor nodes) is given in Table 8.4. By applying
Algorithm 8.1, we find that the FCP with the cost vector [1.55, 1.05, 1.16,
1.05, 1.41, 1.22, 1.41, 1.22, 1.41, 1.28, 1.63, 1.05, 1.16, 1.98, 1.71, 1.16, 1.28,
Figure 8.5 A schematic 1

showing the flow routing
solution for the ten-node
network with base station being 0.8
placed at (0.59, 0.31).
0.6
Y
0.4
0.2
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
X
Table 8.4 Each nodes location, data generation rate, and initial energy for a 20-node
network.
Location Rate Initial energy Location Rate Initial energy
(0.98, 0.49) 0.4 180 (0.09, 0.84) 0.7 60

(0.44, 0.67) 0.8 320 (0.65, 0.62) 0.1 100
(0.57, 0.52) 0.1 340 (0.92, 0.05) 0.1 310
(0.13, 0.19) 0.6 430 (1.00, 0.33) 0.6 280
(0.74, 0.73) 0.1 350 (0.63, 1.00) 0.2 210
(0.24, 0.19) 0.7 310 (0.11, 0.36) 0.3 70
(0.49, 0.38) 0.9 410 (0.89, 0.12) 0.7 420
(0.63, 0.33) 0.7 500 (0.52, 0.86) 0.3 270
(0.76, 0.63) 0.6 270 (0.24, 0.91) 0.9 160
(0.92, 0.33) 0.5 180 (0.40, 0.67) 1.0 180
1.80, 1.05, 1.05] has the maximum network lifetime T = 82.86 among all the
FCPs. Subsequently, we place the base station at a point in the corresponding
subarea, say at point (0.31, 0.79). The corresponding network lifetime is T =
82.91 > T , which is also at least 95% of the optimal value. The flow routing
solution is depicted in Fig. 8.6.
This chapter showed how to design an approximation algorithm to provide

a (1 )-optimal solution to a nonconvex optimization problem. The case
study focused on a classic base station placement problem in a WSN. The
design of the (1 )-approximation algorithm was based on several clever
209 8.11 Problems
Figure 8.6 A schematic 1

showing the flow routing
solution for the 20-node
network with base station being 0.8
placed at (0.31, 0.79).
Y 0.6
0.4
0.2
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
X
techniques, such as discretization of cost parameters (and distances), partition-

ing of the search space into a finite number of subareas, and representation of
subareas with fictitious points (with tight bounds on costs). These three tech-
niques could be exploited to develop approximation algorithms for other prob-
lems. We proved that the approximation algorithm is indeed (1 )-optimal.
The complexity of the algorithm involves solving a polynomial number of LPs,
which is still polynomial.
8.11 Problems
8.1 (a) Explain the difference between constant-factor approximation algo-

rithms and (1 )-optimal (or (1 + )-optimal) approximation algorithms.
(b) Check research literature on wireless networking (from IEEE or ACM
transactions/journals or conference proceedings) and identify one paper for
the constant-factor approximation algorithms.
(c) Based on the paper that you found in Part (b), describe your preference
between the two types of algorithms and give an explanation (in terms of per-
formance and complexity issues).
8.2 Show that the formulation for the optimization problem in Section 8.3.2
is nonconvex.
8.3 Describe the two key techniques used in the design of the approximation
algorithm in this chapter.
8.4 (a) By studying the reference given in the chapter, describe a linear time
algorithm for the smallest enclosing disk problem.
(b) Apply your algorithm of Part (a) to an example having at least five ran-
domly deployed nodes in a two-dimensional area.
8.5 Our discretization procedure in Section 8.6.1 partitions the SED into a
number of irregular subareas. Why is such an irregular subarea division more
attractive than a regular grid-type subarea division?
8.6 Derive (8.10) for Hi .
8.7 What is a fictitious cost point and how is it represented? What is the
difference between a fictitious cost point and a physical point?
8.8 In Definition 8.1, why should we use upper bounds on transmission costs
to define a fictitious cost point?
8.9 Consider the same setting as in 8.1, except that the three nodes 1, 2, and 3
are instead located at (0, 0), (0.93, 0.25), (0.43, 0.75), respectively. Enumerate
all the fictitious cost points in terms of their 3-tuple vectors.
8.10 If the transmission power at each node is upper bounded by a constant
U (U < ), how would you revise Algorithm 8.1?
8.11 Prove that (8.13) is correct.
8.12 In this chapter, we only considered how to optimally locate a single
base station. Suppose now that we have two base stations and we want to
determine where to place each of them optimally so as to maximize the net-
work lifetime. How would you extend Algorithm 8.1 for this two base station
problem? Note that in this case, each sensor can split its data and transmit them
simultaneously to two base stations.
CHAPTER
9 Approximation algorithm
and its applications Part 2
It is not length of life, but depth of life.

Ralph Waldo Emerson
9.1 Introduction
This chapter is a sequel to the last chapter. In the last chapter, we showed
how to design a (1 )-optimal approximation algorithm for a base station
placement problem. The final solution gives a point (or any point within the
corresponding fictitious cost point (FCP) subarea) for the base station. In this
chapter, we extend the base station placement problem to the next level, where
the base station is allowed to move around in the wireless sensor network
(WSN).
The benefits of using a mobile base station to prolong sensor network life-
time are easy to understand. Since the base station is the sink node for data col-
lected by all the sensor nodes in the network, the set of sensor nodes near the
base station would carry a considerable burden in relaying traffic from other
sensor nodes to the base station. By allowing the base station to be mobile,
we could alleviate the traffic burden from this fixed set of sensor nodes to
other sensor nodes in the network, and thus extend network lifetime. Further,
given new advances in unmanned autonomous vehicle (UAV) and customized
robotics, having an unmanned vehicle carrying a base station for sensor data
collection may not be too far from reality.
Although the potential benefit of using a mobile base station to prolong sen-
sor network lifetime is significant, the technical difficulty of this optimization
problem is enormous. There are two components that are tightly coupled in
211
212 Approximation algorithmand its applications Part 2
this problem. First, the location of the base station is time-dependent, i.e., at
different time instances, the base station may be at different locations. Second,
the multi-hop flow routing appears to be dependent on both time and location
of the base station. As a result, an optimization problem with the objective of
maximizing network lifetime needs to consider both base station location and
flow routing, both of which are also time-dependent.
9.2 Case study: The mobile base station problem
In this chapter, we present theoretical results regarding the optimal movement

of a mobile base station. We formulate an optimization problem that incorpo-
rates base station movement and multi-hop flow routing. The solution that we
will present hinges upon two important intermediate results.
The first result shows that as far as the network lifetime objective is con-
cerned, we can transform the time-dependent problem to a location (space)-
dependent problem. In particular, we show that flow routing only depends
on the base station location, regardless of when the base station visits this
location. Further, the specific time instances for the base station to visit a
location is not important, as long as the total sojourn time for the base sta-
tion to be present at this location is the same. This result allows us to focus
on solving a location-dependent problem.
The second result shows that to obtain a (1 )-optimal solution to the
location-dependent problem, we only need to consider a finite set of points
within the smallest enclosing disk (SED) for the mobile base stations loca-
tion. This result follows the same approach in Section 8.6. Then we can find
the optimal sojourn time for the base station to stay at each FCP (as well
as the corresponding flow routing solution) such that the overall network
lifetime (i.e., sum of the sojourn times) is maximized via a single LP prob-
lem. We prove that the proposed solution can guarantee that the achieved
network lifetime is at least (1 ) of the maximum (unknown) network
lifetime.
The rest of this case study is organized as follows. In Section 9.3, we

describe the network model and formally state the base station movement prob-
lem. In Section 9.4, we transform the time-dependent problem to a location-
dependent problem. In Section 9.5, we first develop an optimal solution for
a constrained mobile base station (C-MB) problem, where the base station is
allowed to be present among a set of given locations. Then we present the
solution for the unconstrained mobile base station (U-MB) problem, where
the base station is allowed to roam anywhere in the two-dimensional plane.
Here, we give a formal proof of (1 )-optimality of the proposed algorithm.
213 9.3 Problem and its formulation
In Section 9.6, we present some numerical examples illustrating the efficacy of

the proposed algorithm.
9.3 Problem and its formulation
We consider a set of N sensor nodes deployed over a two-dimensional area,

with the location of each sensor node i N being at a fixed point (xi , yi ). We
assume that each node i generates data at a fixed rate of ri . There is a base
station B for the sensor network and it serves as the sink node for all data
collected by the sensor nodes. Data generated by each sensor node should be
transmitted to the base station via single or multi-hop.
Communication energy is assumed to be the dominant source of energy con-
sumption at a node, which has been modeled in Section 8.3.1. We assume that
each sensor node i N is initially provisioned with an amount of energy ei .
The base station is not constrained with energy and is free to roam in the two-
dimensional plane. Again, network lifetime is defined as the first time instance
when any of the sensor nodes runs out of energy. From (8.1), (8.2), and (8.3), it
is not hard to realize that the location of the base station and the corresponding
multi-hop flow routing among the nodes will affect energy consumption behav-
ior at each node and thus the network lifetime. Table 9.1 lists the notation used
in this chapter.
Our goal is to find how to optimally move a mobile base station to collect
data in a sensor network so that the network lifetime can be maximized. Denote
(x, y)(t) as the position of base station B at time t and T the network lifetime
(which is the objective function of our optimization problem). Then a feasi-
ble flow routing solution realizing this network lifetime T must satisfy both
flow conservation and energy constraint at each sensor node. These constraints
can be formally stated as follows. Denote gij (t) and giB (t) as the data rates
from node i to node j and base station B at time t, respectively. Under multi-
hop multi-path routing, the flow conservation for each node i N at any time
t [0, T ] is
k=i
j =i

gki (t) + ri = gij (t) + giB (t),
kN j N
i.e., for node i, the sum of total incoming flow rates plus self-generated data
rate is equal to the total outgoing flow rates at time t. Note that in our problem,
data generated at each node should be transmitted to the base station in real
time.
Table 9.1 Notation.
General notation
A The movement region for the base station
Cij Transmission energy cost from sensor i to sensor j
ciB (t) Transmission energy cost from sensor i to base station B at time t
ciB (p) Transmission energy cost from sensor i to base station B at point p
ei Initial energy at sensor node i
fij (p) Flow rate from sensor i to sensor j when base station B is at point p
fiB (p) Flow rate from sensor i to base station B when B is at point p
gij (t) Flow rate from sensor i to sensor j at time t
giB (t) Flow rate from sensor i to base station B at time t
N The set of sensor nodes in the network
N = |N |, the number of sensor nodes in the network
p(s) The arrived point when base station B traverses a distance s along P
P The traveling path for base station B
ri Bit rate generated at sensor node i
s(t) Cumulative distance traversed by base station B up to time t
S The total traversed distance at the end of network lifetime
u(s) = v(t1(s)) when the base station traverses point p(s) at time t
U (s) Sojourn time for base station B at distance s
v(t) The
velocity of base station B at time t
w(p) = sZ(p) u(s) when base station B traverses point p but never dwells
W (p) Sojourn time for base station B at point p
(x, y)(t) Location of base station B at time t
(xi , yi ) Location of sensor node i
Z(p) Set of distance s with p(s) = p
Path-loss index, 2 4
1 , 2 Two constant terms in power consumption model for transmission
Power consumption coefficient for receiving data
C-MB problem-specific notation

M The number of predetermined locations
pm The m-th location

TC-MB
The maximum network lifetime achieved by C-MB

C-MB An optimal solution to the C-MB problem
U-MB problem-specific notation

Am The m-th subarea in A
min , C max
CiB iB Lower and upper bounds of ciB (p) for p A
C[h] = 1 (1+)h , the transmission energy cost for the h-th circle
Hi The required number of circles at sensor node i
M The number of subareas under a given
OA , RA The center and radius of A
pm The defined FCP for Am
Sm = {s : p(s) Am , 0 s S}
TU-MB (1 )-optimal network lifetime achieved by U-MB
W (Am ) Sojourn time for base station B in subarea Am
W (pm ) Sojourn time for the base station at FCP pm
Targeted approximation error, > 0 and 1
U-MB A (1 )-optimal solution to the U-MB problem
215 9.4 From time domain to space domain
The energy constraint for each node i N is

7 T k=i j =i

gki (t) + Cij gij (t) + ciB (t) giB (t) dt ei ,
0 kN j N
i.e., total consumed energy due to reception and transmission over time T can-
not exceed its initial energy ei . We have
0
ciB (t) = 1 + 2 (x(t) xi )2 + (y(t) yi )2 ,
by (8.2), where (xi , yi ) is the location of node i.

Denote by A the movement region for the base station, which can be nar-
rowed down to the SED for all nodes in the network, as we have shown in the
last chapter. Note that the SED can be found in polynomial time [167]. The
optimization problem that we are interested in can be formulated as follows:
Maximize T
k=i j =i
subject to kN gki (t) + ri = j N gij (t) + giB (t) (i N , t [0, T ])
8 T !k=i j =i "
0 kN gki (t)+ j N Cij gij (t) + ciB (t) giB (t) dt ei (i N )
!9 "
ciB (t) = 1 + 2 (x(t) xi )2 + (y(t) yi )2 (i N , t [0, T ])
(x, y)(t) A (t [0, T ])
T , gij (t), giB (t) 0 (i, j N , i = j, t [0, T ]),
where the base station location (i.e., (x, y)(t) for t [0, T ]) and the corre-
sponding flow routing (i.e., gij (t) and giB (t) for t [0, T ]) form a joint opti-
mization space for the objective T . Since the left-hand-side in the second
constraint is not a polynomial function of optimization variables, this formula-
tion is in the form of nonpolynomial program.
9.4 From time domain to space domain
The difficulty in the problem formulation lies in that the base station location
(x, y)(t) and flow routing gij (t) and giB (t) are all functions of time. This adds
considerable difficulty in the optimization problem. In this section, we show
that as far as network lifetime performance is concerned, such dependency on
time can be relaxed. Specifically, we will show (Lemma 9.1) that flow routing
only needs to be dependent on the location of the base station and can be
independent of when the base station is present at this location. Further, as
long as the total sojourn time for the base station to be present at this location
is the same, the specific time instance (i.e., when) the base station visits this
location is not important (Lemma 9.2). These results effectively transform the
problem to a location-dependent problem.
We first give the following definition for a time-dependent solution:
Definition 9.1
A time-dependent solution consists of a network lifetime T , a path P =
{(x, y)(t) : t [0, T ]} for the base station, and a flow routing gij (t) and
giB (t) at time t, i, j N , i = j , where (x, y)(t), gij (t) and giB (t) are all
functions of time t.
For such a solution, denote v(t) as the base station velocity8at time t (and
t
thus v(t) is the base station speed at time t). Denote s(t) = =0 v( )d
as the distance traversed by the base station up to time t. Suppose the total
traversed distance at the end of network lifetime T is s(T ) = S. Then we have
s(t) [0, S] for t [0, T ]. Note that for a given path P, we can identify the
corresponding base station location for any s, which we denote as p(s). Denote
U (s) as the sojourn time at distance s. The base station may visit the same point
multiple times. Then multiple distances may correspond to the same point.
Denote Z(p) as the set of such distances that correspond to the same point p.
Then the total sojourn time at a point p is

W (p) = U (s). (9.1)
sZ(p)
Now we give the following definition:
Definition 9.2
A location-dependent solution consists of a path P for the base station,
W (p) at each point p P, flow routing fij (p) and fiB (p), i, j N , i =
j , when the base station is at point p, and a network lifetime T , where
W (p), fij (p), and fiB (p) are all functions of location p.
The following theorem shows that for the objective of network lifetime max-
imization, it is sufficient to consider location-dependent solutions.
Theorem 9.1
The optimal location-dependent solution can achieve the same maximum
network lifetime as the optimal time-dependent solution.
The proof of Theorem 9.1 is based on the following two lemmas:
Lemma 9.1
Given a feasible time-dependent solution, we can construct a location-
dependent solution with the same network lifetime.
Proof. The proof is based on the following construction. For a given time-
dependent solution , it consists of a network lifetime T , a path P for the base
station, and a flow routing gij (t) and giB (t), i, j N , i = j . To construct a
location-dependent solution , we let the base station follow the same path
P and for each point p P, we compute W (p) by (9.1). For , we define
location-dependent flow rates fij (p) and fiB (p) for each point p P by the
average of gij (t) and giB (t) over all visits to p during [0, T ] as follows:
If the base station dwells at p at least once (with W (p) > 0), we define
8 (x,y)(t)=p
t[0,T ] gij (t)dt
fij (p) = , (9.2)
W (p)
8 (x,y)(t)=p
t[0,T ] giB (t)dt
fiB (p) = . (9.3)
W (p)
If the base station traverses p (maybe multiple times) but never dwells, then
there is a unique time corresponding to each s Z(p). Denote such time as
t (s). Define
1
u(s) = , (9.4)
v(t (s))

w(p) = u(s). (9.5)
sZ(p)
Based on u(s) and w(p), we can define

sZ(p) gij (t (s)) u(s)
fij (p) = , (9.6)
w(p)

sZ(p) iB (t (s)) u(s)
g
fiB (p) = . (9.7)
w(p)
To show the data routing scheme with fij (p) and fiB (p) is feasible and
has the same network lifetime T , we need to prove that (i) when the base
station visits each point p P, flow conservation holds at this node, and (ii) at
time T , the energy consumption at each node is the same as that in solution .
Both (i) and (ii) can be intuitively explained by noting that fij (p) and fiB (p)
are defined by the average of gij (t) and giB (t), respectively.
Now we formally prove (i) and (ii).

(i) For flow conservation at point p, if the base station dwells at p at least
once, then we have the following flow conservation:
k=i k=i 8 (x,y)(t)=p 8 (x,y)(t)=p

t[0,T ] gki (t)dt t[0,T ] ri dt
fki (p) + ri = +
W (p) W (p)
kN kN
8 (x,y)(t)=p !k=i "
t[0,T ] kN gki (t) + ri dt
=
W (p)
8 (x,y)(t)=p !j =i "
t[0,T ] j N gij (t) + giB (t) dt
=
W (p)
j =i 8 (x,y)(t)=p 8 (x,y)(t)=p
t[0,T ] gij (t)dt t[0,T ] giB (t)dt
= +
W (p) W (p)
j N
j =i

= fij (p) + fiB (p).
j N
8 (x,y)(t)=p
The first equality holds by (9.2) and the fact that W (p) = t[0,T ] 1dt. The
third equality holds by the flow conservation in solution . The last equality
holds by (9.2) and (9.3).
If the base station traverses (but never dwells at) p, then we have the follow-
ing flow conservation:
k=i k=i
sZ(p) gki (t (s))u(s) sZ(p) ri u(s)
fki (p) + ri = +
w(p) w(p)
kN kN
! k=i "
sZ(p) kN gki (t (s)) + ri u(s)
=
w(p)
! "
k=i
sZ(p) kN gij (t (s)) + giB (t (s)) u(s)
=
w(p)
j =i
sZ(p) gij (t (s))u(s) sZ(p) giB (t (s))u(s)
= +
w(p) w(p)
j N
j =i

j N
The first equality holds by (9.6) and (9.5). The third equality holds by the flow
conservation in solution . The last equality holds by (9.6) and (9.7).
(ii) For energy consumption at time T , we want to show that the energy
consumption at each node i in the constructed location-dependent solution
is the same as that in the given time-dependent solution , i.e.,

k=i U
(s)>0 7 U (s)=0
fki (p(s))U (s) + fki (p(s))u(s)ds
kN s[0,S] s[0,S]

j =i
U
(s)>0 7 U (s)=0
+ Cij fij (p(s))U (s) + fij (p(s))u(s)ds
j N s[0,S] s[0,S]
U
(s)>0 7 U (s)=0
+ ciB (p(s))fiB (p(s))U (s) + ciB (p(s))fiB (p(s))u(s)ds
s[0,S] s[0,S]
k=i
7 T j =i
7 T 7 T
= gki (t)dt + Cij gij (t)dt + ciB (t)giB (t)dt,
kN 0 j N 0 0
where ciB (p) is the energy cost from sensor i to the base station B when the
base station is at point p. To show that the above equality holds, it is sufficient
to show that the following three equalities hold:
U
(s)>0 7 U (s)=0 7 T
fki (p(s))U (s) + fki (p(s))u(s)ds = gki (t)dt
s[0,S] s[0,S] 0
(k, i N , k = i), (9.8)

U
(s)>0 7 U (s)=0 7 T
fij (p(s))U (s) + fij (p(s))u(s)ds = gij (t)dt
s[0,S] s[0,S] 0
(i, j N , j = i), (9.9)

U
(s)>0 7 U (s)=0
ciB (p(s))fiB (p(s))U (s) + ciB (p(s))fiB (p(s))u(s)ds
s[0,S] s[0,S]
7 T
= ciB (t)giB (t)dt (i N ). (9.10)
0
We now prove (9.10). The proofs for (9.8) and (9.9) are very similar (but
simpler) and are left as homework problems.
On the left-hand-side (LHS) of (9.10), we observe that the summation and
integration are over distances in [0, S]. Recall that Z(p) denotes the set of
total traversed distances when the base station visits p P. If the base station
visits p multiple times, then Z(p) has multiple elements. However, for each
s Z(p), ciB (p(s)) is the same since p(s) is the same point p. Further, based
on the definitions in (9.3) and (9.7) for fiB (), we have that fiB (p(s)) is also the
same for each s Z(p). Thus, for a point p P, we can group these distances
in Z(p) together in the summation as well as integration on the LHS of (9.10).
Now we select one distance from each group Z(p), p P, to represent this
group. In particular, we can select the smallest distance in Z(p). Denote Y(P)
as the set of these representatives for each Z(p). Then Y(P) is a subset of
[0, S].
For the summation (first term) on the LHS of (9.10), we have
U
(s)>0
W (p(s))>0 U
(z)>0
ciB (p(s))fiB (p(s))U (s) = ciB (p(z))fiB (p(z))U (z)
s[0,S] sY(P) zZ(p(s))

W (p(s))>0 U
(z)>0
= ciB (p(s))fiB (p(s)) U (z)
sY(P) zZ(p(s))

W (p(s))>0
= ciB (p(s))fiB (p(s)) U (z)
sY(P) zZ(p(s))

W (p(s))>0
= ciB (p(s))fiB (p(s))W (p(s)).
sY(P)
(9.11)
The first equality holds by grouping those distances corresponding to the same
point p(s) that the base station dwells. The second equality holds by ciB (z) =
ciB (s) and fiB (p(z)) = fiB (p(s)) for each z Z(p(s)). The last equality holds
by (9.1).
Following the same token, for the integration (second term) on the LHS of
(9.10), it can be shown that
7 U (s)=0
ciB (p(s))fiB (p(s))u(s)ds
s[0,S]
7 W (p(s))=0
= ciB (p(s))fiB (p(s))w(p(s))ds. (9.12)
sY(P)
Thus by (9.11) and (9.12), we have

W (p(s))>0
LHS of (9.10) = ciB (p(s))fiB (p(s))W (p(s))
sY(P)
7 W (p(s))=0
+ ciB (p(s))fiB (p(s))w(p(s))ds
sY(P)
8 (x,y)(t)=p(s)

W (p(s))>0
t[0,T ] giB (t)dt
= ciB (p(s)) W (p(s))
W (p(s))
sY(P)
7
W (p(s))=0
sZ(p(s)) giB (t (s))u(s)
+ ciB (p(s)) w(p(s))ds
sY(P) w(p(s))

W (p(s))>0 7 (x,y)(t)=p(s)
= ciB (p(s)) giB (t)dt
sY(P) t[0,T ]
7 W (p(s))=0
+ ciB (p(s)) giB (t (s))u(s)ds, (9.13)
sY(P)
sZ(p(s))
where the second equality holds by (9.3) and (9.7).

For the right-hand-side (RHS) of (9.10), we have
7 v(t)>0 7 v(t)=0
RHS of (9.10) = ciB (t)giB (t)dt + ciB (t)giB (t)dt. (9.14)
t[0,T ] t[0,T ]
We now transform each of the above integrations from t-domain to s-domain.
For the case of v(t) = 0 (i.e., the base station dwells at the current
point). Denote t1 (s) as the time when the base station has traversed a dis-
tance s and starts this dwelling period. Denote t2 (s) as the time when
the base station completes this dwelling period. Then U (s) > 0 and for
any time t during [t1 (s), t2 (s)], the base station dwells and v(t) = 0.
Thus,
7 v(t)=0 (s)>0 7 t2 (s)

U
ciB (t)giB (t)dt = ciB (t)giB (t)dt
t[0,T ] s[0,S] t=t1 (s)
U
(s)>0 7 t2 (s)
= ciB (p(s)) giB (t)dt, (9.15)
s[0,S] t=t1 (s)
8 v(t)=0
where the first equality holds as the integration t[0,T ] is limited to those
dwelling periods, each corresponding to a distance s [0, S] with U (s) > 0.
The second equality holds by ciB (t) = ciB (p(s)) for t [t1 (s), t2 (s)].
For the case of v(t) > 0 (i.e., the base station is traversing at the current
point p(s)). We have
7 v(t)>0 7 U (s)=0 1
ciB (t)giB (t)dt = ciB (p(s))giB (s) ds
t[0,T ] s[0,S] v(t (s))
7 U (s)=0
= ciB (p(s))giB (s)u(s)ds, (9.16)
s[0,S]
where giB (s) is the flow rate from sensor i to the base station B when
the base station is traversing at point p(s). The first equality holds by
ds = v(t (s))dt and the second equality holds by (9.4).
Therefore, by (9).14, (9.15), and (9.16), we have
RHS of (9.10)
U
(s)>0 7 t2 (s) 7 U (s)=0
= ciB (p(s)) giB (t)dt + ciB (p(s))giB (s)u(s)ds.
s[0,S] t=t1 (s) s[0,S]
(9.17)
For the summation (first term) on the RHS of (9.17), we have
U
(s)>0 7 t2 (s)
ciB (p(s)) giB (t)dt
s[0,S] t=t1 (s)

W (p(s))>0 U
(z)>0 7 t2 (z)
= ciB (p(z)) giB (t)dt
sY(P) zZ(p(s)) t=t1 (z)

W (p(s))>0 U
(z)>0 7 t2 (z)
= ciB (p(s)) giB (t)dt
sY(P) zZ(p(s)) t=t1 (z)

W (p(s))>0 7 (x,y)(t)=p(s)
= ciB (p(s)) giB (t)dt, (9.18)
sY(P) t[0,T ]
where the first equality holds by grouping those distances corresponding to the
same point p(s) where the base station dwells. The second equality holds by
ciB (p(z)) = ciB (p(s)) for each z Z(p(s)).
Following the same token, for the integration (second term) on the RHS of
(9.17), it can be shown that
7 U (s)=0
ciB (p(s))giB (s)u(s)ds
s[0,S]
7 W (p(s))=0
= ciB (p(s)) giB (t (s))u(s)ds. (9.19)
sY(P)
sZ(p(s))
Therefore, by (9.17), (9.18), and (9.19), we have

W (p(s))>0 7 (x,y)(t)=p(s)
RHS of (9.10) = ciB (p(s)) giB (t)dt
sY(P) t[0,T ]
7 W (p(s))=0
+ ciB (p(s)) giB (t (s))u(s)ds. (9.20)
sY(P)
sZ(p(s))
By (9.13) and (9.20), (9.10) is proved.

Based on our results in (i) and (ii), the constructed location-dependent solu-
tion with P, W (p)(and w(p)), fij (p), and fiB (p) is feasible and has the
same network lifetime T as that achieved by time-dependent solution . This
223 9.5 A (1 )-optimal algorithm
The following lemma further extends Lemma 9.1 and says that the ordering
and specific time instances for the base station to visit a particular point p is
not important.
Lemma 9.2
Under a location-dependent solution, as long as W (p) (and w(p)) at each
point p remains the same, the network lifetime T will remain unchanged
regardless of the ordering and frequency of the base stations presence at
each point.
Lemma 9.2 can be easily proved by analyzing the energy consumption

behavior at each node over time T . We leave its proof as a homework prob-
lem. Combining Lemmas 9.1 and 9.2, and considering the special case that
is optimal, we have Theorem 9.1.
Based on Theorem 9.1, we conclude that as far as the network lifetime is
concerned, it is sufficient for us to study location-dependent solutions. This
result allows us to develop a provable approximation algorithm in the space
domain, which we will present in the following section.
9.5 A (1 )-optimal algorithm
Note that in the location-dependent problem formulation, there are an infinite

number of points in P. In this section, we first consider the case when the base
station is only allowed to be present at a finite set of M positions. We call this
problem the constrained mobile base station (C-MB) problem. Based on this
intermediate result, we then devise a solution to the general problem where
the base station is allowed to roam anywhere on the two-dimensional plane.
We term the latter problem the unconstrained mobile base station (U-MB)
problem.
9.5.1 Optimal sojourn time computation for the C-MB problem

We now show that the C-MB problem can be formulated as an LP, which can
be solved in polynomial time. Recall that in the C-MB problem, the location of
the base station is limited to a finite set of M locations pm , m = 1, 2, . . . , M.
Thus if the base station dwells at pm at least once, then we have W (pm ) > 0
otherwise (i.e., base station never dwells at pm , we have W (pm ) = 0. We
need to find optimal W (pm ) for all points and the corresponding flow rout-
ing fij (pm ) and fiB (pm ) for each point pm with W (pm ) > 0.
When the base station is at point pm , 1 m M, the flow conservation for

node i N is
k=i
j =i

fki (pm ) + ri = fij (pm ) + fiB (pm ). (9.21)
kN j N
The energy constraint for node i N , at the end of network lifetime T is

M k=i
j =i

fki (pm ) + Cij fij (pm ) + ciB (pm ) fiB (pm ) W (pm ) ei .
m=1 kN j N
(9.22)
Note that for each i and pm , ciB (pm ) is a constant.
We can formulate C-MB problem as an LP problem by letting Vij (pm ) =
fij (pm ) W (pm ) and ViB (pm ) = fiB (pm ) W (pm ), where Vij (pm ) (or
ViB (pm )) can be interpreted as the total data volume from sensor node i to
sensor node j (or base station B) when the base station is at pm . We have
LP(C-MB) Maximize T

M
subject to W (pm ) T = 0
m=1
k=i
j =i

Vki (pm ) + ri W (pm ) Vij (pm ) ViB (pm ) = 0
kN j N
(i N , 1 m M) (9.23)

M k=i
j =i

Vki (pm ) + Cij Vij (pm ) + ciB (pm ) ViB (pm ) ei
m=1 kN j N
(i N )
(9.24)
T , W (pm ), Vij (pm ), ViB (pm ) 0 (i, j N , i = j, 1 m M),
where (9.23) and (9.24) follow from (9.21) and (9.22), respectively. Once
we solve the above LP problem, we have W (pm ) for 1 m M. For each
point pm with W (pm ) > 0, we can obtain fij (pm ) and fiB (pm ) by fij (pm ) =
Vij (pm ) ViB (pm )
W (pm ) and fiB (pm ) = W (pm ) . Recall that for those points with W (pm ) = 0,
it means that the base station will not visit those points in this solution.
We summarize the result in this section with the following proposition:
Proposition 9.1
The C-MB problem can be solved via a single LP in polynomial time.
The solution to the above LP problem yields the sojourn time for the base
station at each location pm , m = 1, 2, . . . , M, and the optimal flow routing
when the base station is at pm . So far, we assume that base station B can
move from one point to another in zero time. We will discuss how to relax this
assumption in Section 9.5.4.
9.5.2 Solution to the U-MB problem and proof of (1 )-optimality

We now convert the U-MB problem to a C-MB problem with the (1 ) net-
work lifetime performance guarantee. Our approach is to exploit the energy
cost function and how the location of the base station affects the energy cost.
Note that the location of the base station is embedded in the cost parameter ciB
and that these cost parameters directly affect network lifetime. Thus, to design
a (1 )-optimal algorithm, we divide disk A into subareas (see Section 8.6),
with each subarea to be associated with some nice properties for ciB s that can
be used to prove (1 )-optimality.
Once we have M nonuniform subareas and represent them by M FCPs (see
Section 8.6), we can readily apply the LP approach discussed in Section 9.5.1
to formulate an optimization problem on these M FCPs. In the following, we
will show how to construct a (1 )-optimal solution to the U-MB problem
by solving the C-MB problem on the FCPs.
Denote U-MB as an optimal solution to the U-MB problem and TU-MB as
the maximum network lifetime, both of which are unknown. Our objective
is to find a solution to the U-MB problem that has provable (1 )-optimal

network lifetime. Denote C-MB as an optimal solution to the C-MB problem
obtained by solving an LP problem for the M FCP pm , m = 1, 2, . . . , M, and

TC-MB the corresponding network lifetime.
Our roadmap to construct a solution to the U-MB problem and to prove its
(1 )-optimality is as follows. In Theorem 9.2, we will prove that TC-MB

(1 )TU-MB (see Fig. 9.1). Since the optimal solution C-MB corresponding

to TC-MB is based on the M FCPs instead of physical points, in Theorem 9.3
we will further show how to construct a solution U-MB to the U-MB problem

based on C-MB and prove that the corresponding network lifetime is (1 )-

optimal, i.e., TU-MB (1 )TU-MB (see Fig. 9.1).
Theorem 9.2
For a given > 0, define subareas Am and FCPs pm , m = 1, 2, . . . , M, as

in Section 9.5.2. Then we have TC-MB
(1 ) TU-MB .
Figure 9.1 Comparison of

network lifetimes under
different solutions that are used T*UMB
to construct a (1 )-optimal TUMB
T*CMB
Network lifetime
solution.
(1e)T*UMB
CMB UMB
To prove that Theorem 9.2 is true, we need the following lemma:
Lemma 9.3
Given a feasible solution U-MB to the U-MB problem with a network life-
time TU-MB , we can construct a solution C-MB to the C-MB problem with a
network lifetime TC-MB (1 ) TU-MB .
The proof of this lemma is based on the following construction. Solution

U-MB consists of a specific path P for the base station, W (p) (and w(p)),
fij (p), fiB (p) values for each point p P, and a network lifetime TU-MB .
For subareas Am , m = 1, 2, . . . , M, denote W (Am ) as the total sojourn time
during [0, TU-MB ] when the base station B is within each subarea Am . We have
U
(s)>0 7 U (s)=0
W (Am ) = U (s) + u(s)ds, (9.25)
sSm sSm
where Sm = {s : p(s) Am , 0 s S}. To construct solution C-MB , we

can let the base station spend W (pm ) amount of time on FCP pm , m =
1, 2, . . . , M, where
W (pm ) = (1 ) W (Am ),
and, for each point pm with W (pm ) > 0, set the flow routing when the base
station is at pm as

U(s)>0 7 U (s)=0
1
fij (pm ) = fij (p(s))U (s) + fij (p(s))u(s)ds ,
W (Am ) sSm
sSm

U(s)>0 7 U (s)=0
1
fiB (pm ) = fiB (p(s))U (s) + fiB (p(s))u(s)ds .
W (Am ) sSm
sSm
The details of the proof of Lemma 9.3 are similar to the proof of Lemma 9.1
and are left in a homework problem.
Lemma 9.3 is a powerful result. With this lemma, we are now ready to prove
Theorem 9.2.
Proof. Consider the special case of Lemma 9.3 where the given solution to
the U-MB problem is an optimal solution U-MB
with network lifetime TU-MB .
By Lemma 9.3, we can transform it into a solution to the C-MB problem with

a network lifetime at least (1 )TU-MB , i.e., there is a solution to the C-MB
problem on the FCPs with a network lifetime at least (1 )TU-MB . As a result,

the optimal solution C-MB to the C-MB problem must have a network lifetime

TC-MB
(1 )TU-MB .
Theorem 9.2 guarantees that the network lifetime obtained by the LP solu-
tion on the M FCPs is at least (1 ) of TU-MB . However, a FCP may not be
mapped to a physical point, which is required in the final solution. In the fol-
lowing theorem, we show how to construct a solution with each point being
physically realizable. Further, the network lifetime for this constructed solu-
tion is greater than or equal to the maximum network lifetime for the C-MB

problem, i.e., TU-MB TC-MB . As a result, this new solution is (1 )-optimal.
Theorem 9.3
For a given > 0, define subareas Am and FCPs pm , m = 1, 2, . . . , M,

as discussed in Section 9.5.2. Given an optimal solution C-MB on these

M FCPs with W (pm ), fij (pm ), fiB (pm ), and a network lifetime TC-MB ,a
(1 )-optimal solution U-MB to the U-MB problem can be constructed
by having the base station stay in Am for
W (Am ) = W (pm ) (9.26)
amount of time and by having a corresponding flow routing for any physical
point p Am as
fij (p) = fij (pm ), (9.27)

fiB (p) = fiB (pm ). (9.28)
In Theorem 9.3, note that in the constructed solution to the U-MB problem,
as long as the base station is within Am (any point in this subarea), the flow
routing is the same.
Proof. We will show that the constructed solution U-MB is feasible, i.e., (i)
flow conservation holds at any point, and (ii) the network lifetime of U-MB is at

least TC-MB
. Based on (i) and (ii), as well as the fact that TC-MB
(1 )TU-MB

(see Theorem 9.2), we know that TU-MB must be at least (1 )TU-MB , i.e.,
U-MB is (1 )-optimal.
(i) For flow conservation, when the base station location p is in Am , we have
k=i
k=i

fki (p) + ri = fki (pm ) + ri
kN kN
j =i

= fij (pm ) + fiB (pm )
j N
j =i

j N
The first equality holds by (9.27). The second equality holds by the flow
conservation in solution C-MB . The third equality holds by (9.27) and
(9.28). Thus, solution U-MB is feasible.
(ii) The total energy consumption at node i N by time TC-MB in solution
U-MB is

k=i
U(s)>0 7 U (s)=0
fki (p(s))U (s) + fki (p(s))u(s)ds
kN s[0,S] s[0,S]

j =i
U
(s)=0 7 U (s)=0
+ Cij fij (p(s))U (s) + fij (p(s))u(s)ds
j N s[0,S] s[0,S]
U
(s)>0 7 U (s)=0
+ ciB (p(s))fiB (p(s))U (s)+ ciB (p(s))fiB (p(s))u(s)ds.
s[0,S] s[0,S]
(9.29)
For the last two terms in (9.29), we have
U
(s)>0 7 U (s)=0
ciB (p(s))fiB (p(s))U (s) + ciB (p(s))fiB (p(s))u(s)ds
s[0,S] s[0,S]

M U
(s)>0 7 U (s)=0
= ciB (p(s))fiB (p(s))U (s) + ciB (p(s))fiB (p(s))u(s)ds
m=1 sSm sSm

M U
(s)>0 7 U (s)=0
(p )U (s) +
ciB (pm )fiB (p )u(s)ds
ciB (pm )fiB
m m
m=1 sSm sSm

M U
(s)>0 7 U (s)=0
= (p )
ciB (pm )fiB U (s) + u(s)ds
m
m=1 sSm sSm

M
= (p )W (A )
ciB (pm )fiB m m
m=1
M
= (p )W (p )
ciB (pm )fiB (9.30)
m m
m=1
Algorithm 9.1 A (1 )-optimal algorithm

1. Within A, compute CiB min
and CiBmax
for each node i N by (8.7) and (8.8).
2. For a given > 0, define a sequence of costs C[1], C[2], . . . , C[Hi ] by
(8.9), where Hi is defined by (8.10).
3. For each node i N , draw a sequence of (Hi 1) circles corresponding to
cost C[h] centered at node i, 1 h < Hi . The intersection of these circles
within disk A will divide A into M subareas A1 , A2 , . . . , AM .
4. For each subarea Am , 1 m M, define a FCP pm , which is represented
by an N -tuple cost vector [c1B (pm ), c2B (pm ), . . . , cNB (pm )], where ciB (pm )
is defined by (8.12).
5. For the C-MB problem on these M FCPs, apply the LP formulation in Sec-

tion 9.5.1 and obtain an optimal solution C-MB with W (pm ), fij (pm ), and

fiB (pm ).
6. Construct a (1 )-optimal solution U-MB to the U-MB problem based on

C-MB using the procedure in Theorem 9.3.
for i N . The second inequality holds by ciB (p(s)) ciB (pm ) in Prop-
erty 8.1 and (9.28). The fourth equality holds by (9.25). The last equality
holds by (9.26).
Following the same token, it can be shown that

U
(s)>0 7 U (s)=0
M
fij (p(s))U (s) + fij (p(s))u(s)ds = fij (pm )W (pm ).
s[0,S] s[0,S] m=1
(9.31)
for i, j N and i = j . Thus, by (9.30) and (9.31), the total energy consump-

tion at node i by time TC-MB , which is shown in (9.29), is no more than
k=i

M j =i

M
fki (pm )W (pm ) + Cij fij (pm )W (pm )
kN m=1 j N m=1

M
+ ciB (pm )fiB (pm )W (pm ) ei ,
m=1

where the inequality holds by the energy constraint in solution C-MB . Thus,

the network lifetime of solution U-MB is at least TC-MB (1 )TU-MB . This
9.5.3 Summary of algorithm and example

We now summarize the design of the (1 )-optimal algorithm in Algo-
rithm 9.1. In Algorithm 9.1, Step 5 has the highest complexity (solving an
LP problem) among all steps. Since there are (Hi 1) circles radiating from
Table 9.2 Sensor location, data rate, and initial energy of the example
sensor network.
Node index (xi , yi ) ri ei
1 (0.2, 0.9) 0.6 170

2 (0.4, 0.6) 1.0 420
3 (0.6, 0.3) 0.8 460
4 (1.0, 0.2) 0.4 230
Figure 9.2 The SED for the

Y
example sensor network.
1
1
0.8
2
0.6
OA
RA
0.4
3
4
0.2
0
0 0.2 0.4 0.6 0.8 1 X
sensor node i N and one circle for disk A, the total number of subareas
M obtained through the intersection of these circles is upper bounded by

O([1 + N i=1 (Hi 1)] ) = O((N/) ). Thus, the LP problem in Step 5 has
2 2
polynomial size and the complexity of the overall algorithm is polynomial.
Example 9.1
To illustrate the steps in Algorithm 9.1, we solve a small four-node WSN
problem as an example. The location, data rate, and initial energy for each
sensor node are shown in Table 9.2, where the units of distance, rate, and
energy are all normalized with appropriate dimensions. We use = 2 in
this example and set 1 = 1, 2 = 0.5 and = 1 under normalized units.
For illustration, we set = 0.2.1
In Step 1, we first identify the SED A with origin OA = (0.60, 0.55) and
radius RA = 0.53 (see Fig. 9.2). Then we have D1,OA = 0.53, D2,OA =
0.21, D3,OA = 0.25, and D4,OA = 0.53. We then find the lower and upper
bounds of ciB for each node i as follows:
min
CiB = 1 = 1,
max
CiB = 1 + 2 (Di,OA + RA ) .
Thus, we have
max
C1B = 1 + 0.5 (0.53 + 0.53)2 = 1.56,
max
C2B = 1 + 0.5 (0.21 + 0.53)2 = 1.27,
max
C3B = 1 + 0.5 (0.25 + 0.53)2 = 1.30,
max
C4B = 1 + 0.5 (0.53 + 0.53)2 = 1.56.
In Step 2, for = 0.2, since
2 2 4
ln(1 + 1 (Di,OA + RA ) )
Hi = ,
ln(1 + )
we have
2 4
ln(1 + 1 (0.53 + 0.53) )
0.5 2
H1 = = 3,
ln(1 + 0.2)
2 4
ln(1 + 1 (0.21 + 0.53) )
0.5 2
H2 = = 2,
ln(1 + 0.2)
2 4
ln(1 + 1 (0.25 + 0.53) )
0.5 2
H3 = = 2,
ln(1 + 0.2)
2 4
ln(1 + 1 (0.53 + 0.53) )
0.5 2
H4 = = 3,
ln(1 + 0.2)
and
C[1] = 1 (1 + ) = 1 (1 + 0.2) = 1.20,
C[2] = 1 (1 + )2 = 1 (1 + 0.2)2 = 1.44,
C[3] = 1 (1 + )3 = 1 (1 + 0.2)3 = 1.73.
In Step 3, we draw a sequence of circles centered at each node i, 1 i
4, and with cost C[h], 1 h < Hi , to divide the SED A into 16 subareas
A1 , A2 , . . . , A16 (see Fig. 9.3).
In Step 4, we define a FCP pm for each subarea Am , 1
m 16. For example, for FCP p1 , we define the 4-tuple cost vec-
tor as [c1B (pm ), c2B (pm ), c3B (pm ), c4B (pm )] = [C[1], C[1], C[2], C[3]] =
[1.20, 1.20, 1.44, 1.73].
In Step 5, we obtain an optimal solution C-MB to C-MB problem on
these 16 FCPs by the LP approach discussed in Section 9.5.1. We obtain the
network lifetime TC-MB = 247.76, W (p7 ) = 144.22, W (p12 ) = 82.50,
W (p16 ) = 21.04, and for all other 13 FCPs, we have W (pm ) = 0 (mean-

ing the base station will not visit these 13 subareas). When the base station
is at FCP p7 , the routing is f1B (p ) = 0.60, f (p ) = 0.51, f (p ) =
7 2B 7 23 7

0.49, f3B (p7 ) = 1.29, and f4B (p7 ) = 0.40. When the base station is at FCP
p12 , the routing is f12 (p ) = 0.60, f (p ) = 1.60, f (p ) = 0.80,
12 2B 12 3B 12

and f4B (p12 ) = 0.40. When the base station is at FCP p16 , the routing is
(p ) = 0.60, f (p ) = 1.60, f (p ) = 2.40, and f (p ) = 2.80.
f12 16 23 16 34 16 4B 16
1
3
1 5
9
6
0.8 10
11
15
13
2
0.6 4
7
0.4
12 16
0.2
8
14
0
0 0.2 0.4 0.6 0.8 1 X
Figure 9.3 The subareas for the example sensor network.
In Step 6, we obtain a (1 )-optimal solution U-MB to U-MB prob-

lem as follows. Let the base station stay at any point in subarea A7
for 144.22 unit of time, stay at any point in subarea A12 for 82.50
unit of time, and stay at any point in subarea A16 for 21.04 unit of
time. When the base station is at a point p in subarea A7 , the rout-
ing is f1B (p) = 0.60, f2B (p) = 0.51, f23 (p) = 0.49, f3B (p) = 1.29, and
f4B (p) = 0.40. When the base station is at a point p in subarea A12 ,
the routing is f12 (p) = 0.60, f2B (p) = 1.60, f3B (p) = 0.80, and f4B (p) =
0.40. When the base station is at a point p in subarea A16 , the rout-
ing is f12 (p) = 0.60, f23 (p) = 1.60, f34 (p) = 2.40, and f4B (p) = 2.80.
The network lifetime for U-MB is greater than or equal to 247.76 and is
(1 )-optimal.
9.5.4 Discussions
We now discuss the design of a path P based on W (pm ) values. Such a
path is certainly not unique. In 9.1, the base station can move from sub-
Table 9.3 Each nodes location, data generation rate and initial energy for a ten-node
network.
Location Data rate Initial energy Location Data rate Initial energy
(0.0, 0.8) 0.8 150 (0.6, 0.7) 0.6 370

(1.0, 1.0) 1.0 200 (0.4, 0.6) 0.2 420
(0.3, 0.4) 0.6 130 (0.2, 0.9) 0.6 100
(0.8, 0.5) 0.8 460 (0.9, 0.1) 0.4 80
(0.7, 0.3) 0.3 170 (0.5, 0.2) 1.0 150
area 7 to 12 and to 16 (denote as (A7 , A12 , A16 )) or, another path can be
(A16 , A12 , A7 ). Note that for any path, as long as the total sojourn time at
each subarea Am is W (pm ), the achieved network lifetime is (1 )-optimal.
Thus, all of these paths are equally good under network lifetime objective.
It may be arguable that one path is better than another under some other
objective, e.g., minimizing the total traveled distance. However, such objec-
tive can be formulated as a separate problem and its discussion is beyond the
scope of this chapter.
Along a path P, it is possible that one subarea and the next subarea that the
base station visits are not adjacent. We argue that the traveling time between
two subareas (e.g., minutes) is likely on a much smaller time scale than net-
work lifetime (e.g., months). It can be shown that if buffering is available at
sensor nodes when base station is in transition from one subarea to the next
subarea, then the (1 )-optimal network lifetime can still be achieved. In
this case, a node only needs to slightly delay its transmission until the base
station arrives at the next subarea and then empties the buffer with a higher
rate for a brief period of time.
Now we apply the (1 )-optimal algorithm for different sized networks

and use numerical results to demonstrate the efficacy of the algorithm. We
consider four randomly generated networks consisting of ten, 20, 50, and
100 nodes deployed over a unit square area, respectively. The data rate and
initial energy for each node are randomly generated between [0.1, 1] and
[50, 500], respectively. The units of distance, rate, and energy are all normal-
ized appropriately. The normalized parameters in energy consumption model
are 1 = 2 = = 1. We assume the path-loss index = 2 and set = 0.05.
The network setting (location, data rate, and initial energy for each node)
for the ten-node network is given in Table 9.3. By applying Algorithm 9.1, we
obtain a (1 )-optimal network lifetime 142.86, which is guaranteed to be
at least 95% of the optimum. In Table 9.4, we have seven subareas that will
Table 9.4 Sojourn time at

each optimal location for
the ten-node network.
Am (x, y) W (Am )
(0.93, 0.96) 4.28

(0.88, 0.19) 3.54
(0.68, 0.90) 0.04
(0.94, 0.89) 50.36
(0.69, 0.22) 44.00
(0.90, 0.40) 27.15
(0.23, 0.86) 13.49
Figure 9.4 Network topology 1

and optimal locations for base
station movement for the
0.8
ten-node network.
0.6
Y
0.4
0.2
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
X
be visited by the base station in the (1 )-optimal solution (also shown in

Fig. 9.4). For illustration purpose, we use a star to represent the corresponding
subarea that the base station will visit in the solution. For example, we put a
star on location (0.93, 0.96) to represent the subarea that contains this point.
Table 9.4 also lists the corresponding sojourn time for the base station to stay in
each of these seven subareas. The flow routing solution when the base station is
in each of the seven subareas is different, as expected. Fig. 9.5 shows a possible
path for the ten-node network. Note that as we discussed in Section 9.5.4, such
a path is not unique.
It is worth noting that for 95% optimality, only seven subareas need to be
visited by the base station. It turns out that for 20-, 50-, and 100-node networks,
the number of subareas that needs to be visited by the base station is also very
small (six subareas for 20-node network, eight subareas for 50-node network,
and 12 subareas for 100-node network). This observation is not obvious. But it
is a good news as it hints that the base station may not need to move frequently
to many different locations to achieve near-optimal solution.
Table 9.5 Each nodes location, data generation rate and initial energy for a 20-node
network.
(0.52, 0.02) 0.6 480 (0.29, 0.14) 0.6 120
(0.74, 0.76) 0.3 310 (0.05, 0.99) 0.4 60
(0.95, 0.03) 0.8 150 (0.84, 0.06) 1.0 180
(0.53, 0.63) 0.6 220 (0.99, 0.37) 0.4 340
(0.58, 1.00) 0.4 230 (0.73, 0.67) 0.8 220
(0.48, 0.84) 0.7 160 (0.53, 0.27) 0.5 380
(0.17, 0.83) 0.1 380 (0.57, 0.05) 0.7 250
(0.73, 0.39) 0.1 500 (0.88, 0.84) 0.2 240
(0.36, 0.98) 0.1 430 (0.26, 0.12) 0.9 440
(0.76, 0.02) 0.7 500 (0.71, 0.21) 0.3 70
Figure 9.5 A possible base 1

station moving path for the
ten-node network.
0.8
0.6
Y
0.4
0.2
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
X
The network setting for a small 20-node network (with location, data rate,
and initial energy for each of the 20 sensor nodes) is given in Table 9.5. By
applying Algorithm 9.1, we obtain a (1 )-optimal network lifetime 144.23.
Again, we use a star to represent the subarea that base station will visit in
the solution. For this particular 20-node network setting, we have six subareas
(see Fig. 9.6) that the base station will visit in the final solution, with the cor-
responding sojourn time in each subarea shown in Table 9.6. Again, we show
a possible path for the 20-node network in Fig. 9.7.
The network setting for the 50-node network (with location, data rate, and
initial energy for each of the 50 sensor nodes) is given in Table 9.8. By apply-
ing Algorithm 9.1, we obtain a (1 )-optimal network lifetime 122.30. In
Table 9.7, we have eight subareas (see Fig. 9.8) that the base station will visit
in the (1 )-optimal solution, as well as the sojourn time for the base station
Table 9.6 Sojourn time at each

optimal location for the 20-node
network.
Am (x, y) W (Am )
(0.28, 0.33) 2.86

(0.27, 0.47) 8.44
(0.95, 0.86) 9.62
(0.80, 0.06) 5.34
(0.70, 0.11) 108.05
(0.88, 0.05) 9.92

20-node network. 0.8
0.6
Y
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1
X

20-node network.
0.8
0.6
Y
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1
X
in each of these eight subareas. Fig. 9.9 shows a possible path for the 50-node
network.
Finally, we consider a 100-node network shown in Table 9.9. We omit
to list the each nodes coordinates, data rate, and initial energy to conserve

optimal location for the 50-node
network.
Am (x, y) W (Am )
(0.13, 0.11) 0.39

(0.17, 0.64) 1.47
(0.32, 0.90) 26.23
(0.25, 0.61) 27.72
(0.49, 0.12) 3.43
(0.94, 0.10) 9.66
(0.34, 0.26) 8.37
(0.16, 0.30) 45.03

0.6
Y
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1
X

50-node network.
0.8
0.6
Y
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1
X
Table 9.8 Each nodes location, data generation rate and initial energy for a 50-node
network.
(0.52, 0.24) 1.0 290 (0.80, 0.63) 0.8 260

(0.68, 0.91) 0.8 160 (0.00, 0.59) 0.3 50
(0.32, 0.90) 0.4 480 (0.81, 0.54) 0.7 150
(0.54, 0.55) 0.1 500 (0.78, 0.46) 0.8 150
(0.78, 0.08) 0.1 140 (0.84, 0.18) 0.7 160
(0.94, 0.02) 0.8 300 (0.61, 0.43) 0.7 400
(0.06, 0.12) 0.1 220 (0.11, 0.10) 0.9 300
(0.45, 0.56) 0.4 370 (0.20, 0.87) 0.5 470
(0.83, 0.29) 0.4 400 (0.16, 0.05) 0.9 140
(0.17, 0.52) 0.7 160 (0.61, 0.42) 0.2 450
(0.42, 0.98) 0.8 280 (0.72, 0.84) 0.5 480
(0.55, 0.88) 0.4 320 (0.56, 0.82) 0.9 240
(0.95, 0.43) 0.6 390 (0.42, 0.90) 0.3 490
(0.99, 0.41) 0.8 180 (0.37, 0.71) 0.1 470
(0.16, 0.53) 0.8 190 (0.71, 0.70) 0.6 220
(0.89, 0.62) 0.7 340 (0.47, 0.97) 0.6 140
(0.69, 0.52) 0.7 220 (0.16, 0.73) 0.1 150
(0.86, 0.79) 0.4 50 (0.51, 0.47) 0.1 90
(0.23, 0.75) 0.6 150 (0.77, 0.63) 1.0 390
(0.43, 0.99) 0.5 290 (0.65, 0.18) 0.4 340
(0.60, 0.96) 0.3 500 (0.48, 0.19) 0.5 70
(0.56, 0.62) 0.4 420 (0.09, 0.66) 0.8 140
(0.50, 0.68) 1.0 170 (0.48, 0.62) 0.6 300
(0.17, 0.66) 1.0 250 (0.93, 0.36) 0.5 270
(0.66, 0.30) 0.1 100 (0.28, 0.19) 0.8 160
Table 9.9 Each nodes location, data generation rate and initial energy for a 100-
node network.
(0.85, 0.13) 0.1 160 (0.43, 0.44) 0.5 290

(0.74, 0.99) 0.5 250 (0.98, 0.16) 0.8 160
(0.31, 0.06) 0.9 430 (0.27, 0.71) 0.1 190
(0.93, 0.67) 0.7 240 (0.68, 0.09) 0.2 130
(0.65, 0.56) 0.3 310 (0.77, 0.77) 0.2 320
(0.75, 0.12) 0.4 90 (0.49, 0.44) 0.4 400
(0.63, 0.74) 0.9 470 (0.35, 0.83) 0.5 190
(0.88, 0.17) 0.9 440 (0.88, 0.12) 0.4 460
(0.91, 0.48) 0.1 180 (0.08, 0.27) 0.8 210
(0.94, 0.02) 0.9 220 (0.77, 0.72) 0.1 380
(0.26, 0.35) 0.8 150 (0.20, 0.74) 1.0 470
Table 9.9 (Cont.)
(0.77, 0.35) 0.5 160 (0.20, 0.56) 0.7 320

(0.42, 0.73) 0.3 100 (0.66, 0.12) 1.0 190
(0.65, 0.58) 0.5 320 (0.44, 0.06) 0.5 450
(0.93, 0.15) 0.7 150 (0.61, 0.05) 0.7 210
(0.32, 0.04) 0.2 230 (0.10, 0.26) 0.7 480
(0.63, 0.96) 0.6 110 (0.52, 0.85) 0.1 440
(0.51, 0.75) 0.5 360 (0.06, 0.65) 0.4 320
(0.03, 0.27) 0.1 350 (0.49, 0.16) 0.8 380
(0.64, 0.43) 0.1 360 (0.79, 0.15) 0.4 480
(0.67, 0.40) 0.6 230 (0.27, 0.50) 0.4 140
(0.84, 0.38) 0.1 210 (0.22, 0.81) 0.6 290
(0.21, 0.16) 0.7 430 (0.77, 0.46) 0.6 370
(0.37, 0.93) 0.9 460 (0.06, 0.04) 0.8 230
(0.86, 0.96) 0.1 90 (0.95, 0.95) 0.3 80
(0.77, 0.48) 0.6 270 (0.10, 0.42) 0.4 250
(0.56, 0.09) 0.9 270 (0.91, 0.86) 0.7 280
(0.72, 0.26) 1.0 260 (0.33, 0.55) 0.7 490
(0.88, 0.51) 0.3 110 (0.38, 0.33) 0.1 490
(0.66, 0.49) 0.1 90 (0.01, 0.88) 0.9 420
(0.80, 0.92) 0.8 170 (0.07, 0.19) 0.3 230
(0.48, 0.78) 0.1 70 (0.12, 0.31) 0.9 470
(0.87, 0.38) 1.0 150 (0.95, 0.72) 0.8 370
(0.92, 0.68) 0.1 60 (0.85, 0.65) 1.0 210
(0.71, 0.12) 0.8 50 (0.51, 0.94) 0.2 170
(0.69, 0.79) 0.1 200 (0.24, 0.78) 0.6 250
(0.38, 0.90) 0.2 180 (0.26, 0.48) 0.2 70
(0.77, 1.00) 0.2 120 (0.87, 0.74) 0.4 150
(0.44, 0.34) 0.9 200 (0.80, 0.47) 0.6 440
(0.54, 0.21) 1.0 340 (0.71, 0.06) 0.8 380
(0.44, 0.48) 0.1 450 (0.13, 0.68) 0.2 460
(0.54, 0.05) 0.6 130 (0.60, 0.98) 0.2 460
(0.91, 0.04) 0.7 220 (0.89, 0.01) 0.7 430
(0.57, 0.21) 0.2 220 (0.95, 0.33) 0.2 430
(0.61, 0.77) 0.4 160 (0.09, 0.09) 0.3 260
(0.75, 0.55) 0.2 170 (0.33, 0.86) 0.5 490
(0.39, 0.41) 0.8 190 (0.01, 0.72) 0.1 170
(0.26, 0.32) 0.8 230 (0.15, 0.42) 0.1 160
(0.25, 0.97) 0.9 310 (0.03, 0.69) 0.5 190
(0.35, 0.92) 0.3 290 (0.01, 0.27) 0.6 120
space. They are all randomly generated as we described early in this sec-
tion. By applying Algorithm 9.1, we obtain a (1 )-optimal network lifetime
149.45. For this particular 100-node network setting, we have 12 subareas (see
Fig. 9.10) that the base station will visit, with the corresponding sojourn time in
each subarea shown in Table 9.10. Fig. 9.11 shows a possible path for the 100-
node network.

optimal location for the 100-
node network.
Am (x, y) W (Am )
(0.75, 0.83) 0.15

(0.86, 0.05) 0.38
(0.69, 0.28) 50.14
(0.24, 0.04) 2.10
(0.59, 0.88) 0.57
(0.81, 0.82) 0.09
(0.20, 0.53) 21.21
(0.91, 0.09) 0.05
(0.40, 0.68) 41.64
(0.89, 0.19) 3.16
(0.61, 0.79) 23.82
(0.19, 0.02) 6.14

0.6
Y
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1
X
9.7 Chapter summary
This chapter is a sequel to the last chapter. Again, our interest is in the design
of a (1 )-optimal approximation algorithm. But the problem is much harder
than that in the last chapter. By allowing the base station to be mobile, both the
location of the base station and the multi-hop flow routing in the network are
time-dependent.
To address this problem, we showed that as far as network lifetime objec-
tive is concerned, we can transform the time-dependent problem to a location
(space)-dependent problem. In particular, we showed that flow routing only
depends on the base station location, regardless of when the base station visits
this location. Further, the specific time instances for the base station to visit a
241 9.8 Problems

100-node network.
0.8
0.6
Y
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1
X
location are not important, as long as the total sojourn time for the base station
to be present at this location is the same. This result allowed us to focus on
solving a location-dependent problem.
Based on the above result, we further showed that to obtain a (1 )-
optimal solution to the location-dependent problem, we only need to consider
a finite set of points within the SED for the mobile base stations location.
Here, we followed the same approach as that in Section 8.6, i.e., discretization
of energy cost through a geometric sequence, division of a disk into a finite
number of subareas, and representation of each subarea with a FCP. Then we
can find the optimal sojourn time for the base station to stay at each FCP (as
well as the corresponding flow routing solution) so that the overall network
lifetime (i.e., sum of the sojourn times) is maximized via a single LP problem.
We proved that the proposed solution can guarantee that the achieved network
lifetime is at least (1 ) of the maximum (unknown) network lifetime.
This chapter offers some excellent examples on how to transform a prob-
lem from time domain to space domain and how to prove results through
construction. Students are encouraged to gain a deep understanding of these
techniques, which should transcend to other optimization problems in wireless
networks.
9.8 Problems
9.1 What is the benefit of using mobile base station? Compared to the static
base station problem in the last chapter, what are the challenges in employing
a mobile base station?
9.2 Show that the formulation for the optimization problem in Section 9.3 is
nonconvex.
9.3 For the optimization problem formulated in Section 9.3, can we dis-
cretize time t to obtain an approximation algorithm?
9.4 Describe the three key techniques used in the design of the approxima-
tion algorithm in the chapter.
9.5 What is the relationship between a time-dependent solution and a
location-dependent solution?
9.6 Can we use Eqs. (9.2) and (9.3) to define fij (p) and fiB (p) for every
point p P? Why? Can we use Eqs. (9.6) and (9.7) to define fij (p) and
fiB (p) for every point p P? Why?
9.7 Prove Eq. (9.8).
9.8 Prove Lemma 9.2.
9.9 Complete the proof of Lemma 9.3.
9.10 In Section 9.5.4, it is said that the design of a path P based on W (pm )
values is not unique. We may use an additional objective to select the best path,
e.g., minimizing the total traveled distance. Discuss other possible objective(s)
for path selection.
PART
III Methods for Efficient Heuristic

Solutions
CHAPTER
10 An efficient technique for

mixed-integer optimization
Remember, happiness doesnt depend upon who you are or what you have; it
depends solely upon what you think.
Dale Carnegie
10.1 Sequential fixing: an introduction
In Chapter 5, we presented a branch-and-bound framework for solving mixed-

integer nonlinear programs (MINLPs), which can be used to solve a broad
class of problems. Furthermore, it can guarantee a (1 )-optimal solution
for any given 0 but its worst-case complexity is exponential. In this chap-
ter, we present another technique called Sequential Fixing (SF) to solve certain
mixed-integer optimization problems. SF is designed to iteratively determine
(fix) binary integer variables. Unlike branch-and-bound, it is a heuristic pro-
cedure and has polynomial-time complexity. Nevertheless, it is a very efficient
technique, and, in a number of problem instances, we found that this technique
can offer highly competitive solutions.
The basic idea behind SF is as follows. For some mixed-integer optimization
problems, such as certain MINLP and any mixed-integer linear programming
(MILP), if we were able to set the optimal values for all the integer vari-
ables and thereby reduce the original problem to an LP, then we can solve
the reduced problem optimally in polynomial time. Thus, the key challenge in
such contexts is how to determine the values for all the integer variables. This
can be done by examining the linear or continuous relaxation of the original
problem, which is obtained by relaxing all the integer variables to continu-
ous variables. Although the solution to this relaxation may not have an inte-
ger value for each integer variable, we can set (i.e., fix) the values of one or
245
246 An efficient technique for mixed-integer optimization
more integer variables based on the closeness in the relaxed solution to certain
integer values. Instead of determining all integer variable values via a single
relaxation, we can fix only one or a few integer variables in each iteration. For
the remaining (unfixed) integer variables, we can solve a new relaxation (with
some integer variables values being already fixed) and then fix one or more
integer variables. This SF procedure terminates after we fix all the integer vari-
ables. The values of other variables in the original problem can be obtained by
solving the resulting LP problem.
Unlike most techniques in other chapters, SF is a heuristic procedure. To
measure its performance, we can compare its solution value to some perfor-
mance bound, e.g., a lower bound for a minimization problem, or an upper
bound for a maximization problem. Such a bound can be obtained via the ini-
tial relaxation to the original problem. Note that the optimal objective value lies
between this bound and the solution obtained by the SF algorithm. Therefore,
if we can show (e.g., via a large number of simulations) that the SF solution
has a value very close to the computed relaxation-based bound, then we can
claim that the SF solution must be even closer to the optimum, thus validating
its performance.
In the case study of this chapter, we show how SF can be employed to solve
a spectrum sharing problem in a multi-hop cognitive radio network (CRN).
10.2 Case study: Spectrum sharing for cognitive radio networks
Consider a multi-hop CRN. For such a network, each node senses a set of
spectrum bands that it can use for communication. Due to the unequal size
of spectrum bands, it may be necessary to further divide a large band into
subbands (likely of unequal size) to schedule transmission and reception.
There are many fundamental problems that can be posed for such a wireless
network in the context of rates and capacity. In this chapter, we consider the
following problem. Suppose there exists a set of user sessions in the network
that is characterized by several sourcedestination pairs, each having a certain
rate requirement. Then, how can we perform spectrum allocation, scheduling
and interference avoidance, and multi-hop flow routing such that the required
network-wide radio spectrum resource is minimized?
To formulate the problem mathematically, we characterize the structure
and constraints of a CRN at multiple layers. Special attention is given to
the modeling of spectrum sharing and unequal (nonuniform) subband divi-
sion, scheduling and interference modeling, and multi-hop routing. We for-
mulate an optimization problem with the objective of minimizing the required
network-wide radio spectrum resource for a set of session rate requirements.
Since such a problem formulation is a mixed-integer nonlinear programming
(MINLP), which is NP-hard in general [46], we aim to derive a near-optimal
solution.
247 10.3 Mathematical modeling and problem formulation
We present a near-optimal algorithm for the formulated MINLP problem.

First, we develop a lower bound for the objective function by relaxing the
integer variables and employing a linearization technique. This lower bound
will be used as a measure for the quality of any solution. Then we present an SF
solution procedure where the determination of integer variables is performed
iteratively through a sequence of LPs. After fixing all the integer variables,
the other variables in the optimization problem can be derived by solving the
residual LP problem. Since the solution obtained by the proposed SF algorithm
represents an upper bound for the minimization problem, we compare it to the
lower bound developed earlier. Simulations show that the results obtained by
the SF algorithm are very close to the lower bound, thus suggesting that (1) the
lower bound is very tight; and (2) the solution obtained by the SF algorithm is
very close to the optimum, and thus is near-optimal.
characterize a CRN based on multiple layers and formulate its structure in
terms of mathematical constraints. We also describe the optimal spectrum
sharing problem and formulate it as an MINLP problem. In Section 10.4,
we develop a lower bound for this MINLP problem by relaxing integer vari-
ables and using linearization. In Section 10.5, we describe an SF algorithm.
Section 10.6 presents simulation results and demonstrates the near-optimal
performance of the SF algorithm. Section 10.7 summarizes this chapter.
10.3 Mathematical modeling and problem formulation
We consider an ad hoc network consisting of a set N of nodes. There is a set

L of unicast communication sessions. Denote s(l) and d(l) as the source and
destination nodes of session l L, and r(l) as the rate requirement (in b/s) of
session l. Table 10.1 lists the notation used in this chapter.
10.3.1 Modeling of multi-layer characteristics

Spectrum sharing and subband division In a multi-hop CRN, the available
spectrum bands at one node may be different from another node in the network.
Given a set of available frequency bands at a node, the size (or bandwidth) of
each band may differ drastically. For example, among the least-utilized spec-
trum bands found in [106], the bandwidth between [1240, 1300] MHz (allo-
cated to amateur radio) is 60 MHz, while the bandwidth between [1525, 1710]
MHz (allocated to mobile satellites, GPS systems, and meteorological applica-
tions) is 185 MHz. Such a large difference in bandwidths among the available
bands suggests the need to further divide a large band into smaller subbands
for a more flexible and efficient frequency allocation. Since an equal subband
division of the available spectrum band is likely to yield suboptimal perfor-
mance, an unequal division may be necessary.
Table 10.1 Notation.
Symbol Definition
d(l) Destination node of session l

dij Distance between nodes i and j
fij (l) Data rate that is attributed to session l on link i j
gij Propagation gain from node i to node j
Ijm The set of nodes that can use band m and are within the interference
range of node j
K (m) The maximum number for subband divisions in band m
L The set of active user sessions in the network
Mi Theset of available bands at node i N
M = iN Mi , the set of available bands in the network
M = |M|,(the number of available bands in the network
Mij = Mi Mj , the set of available bands for link i j
r(l) Rate of session l L
RT , RI Transmission range and interference range, respectively
s(l) Source node of session l
Tim The set of nodes that can use band m and are within the transmission
range
of node i
Ti = mMi Tim , the set of nodes within the transmission range of
node i
u(m,k) The fraction of bandwidth for the k-th subband in band m
W (m) Bandwidth of band m M
(m,k)
xij Binary indicator to mark whether or not subband (m, k) is used
for link i j .
Path-loss index
Transmission power spectral density at a transmitter
T The minimum threshold of power spectral density to decode a
transmission at a receiver
I The maximum threshold of power spectral density for interference to
be negligible at a receiver
More formally, we model the union of the available frequency bands among
all the nodes in the network as a set of M unequally sized bands (see Fig. 10.1).
Denote M as the set of these bands and let Mi M be the set of available
bands at node i N , which is likely to be different from that at another node,
say j N , i.e., possibly, Mi = Mj . For example, at node i, Mi may consist
of bands 1, 3, and 5, whereas at node j , Mj may consist of bands 1, 4, and
6. Denote W (m) as the bandwidth of band m M. For more flexible and effi-
cient bandwidth allocation and to overcome the disparity in the bandwidth size
among the spectrum bands, we assume that band m can be further divided into
up to K (m) subbands, each of which may be of unequal bandwidth. Denote
Figure 10.1 A schematic (1)

W W (2) W (m) W (M)
illustration of bands and
subbands in spectrum sharing. 1 2 m M
(m)
u(m,1)W (m) u(m,2)W (m) u(m,K )
W (m)
(m)
(m,1) (m,2) (m,K )
u(m,k) as the fraction of bandwidth for the k-th subband in band m, which is
part of our set of optimization decision variables. Then we have
(m)

K
u(m,k) = 1.
k=1
Note that some u(m,k) -variables can be 0 in the final optimization solution,
in which case we will have a fewer number of subbands than K (m) . As an
example, Fig. 10.1 shows M bands in the network, and for a specific band m,
it displays a further division into K (m) subbands. Thus, the M bands in the

network are effectively divided into M m=1 K
(m) subbands, each of which may
be of a different size.
Transmission range and interference range We assume that the power
spectral density from the transmitter of a CR node is . In this chapter, we
assume that all nodes use the same power density for transmission. A widely
used model for power propagation gain is [56]
gij = dij , (10.1)
where is an antenna related constant, is the path-loss index, and dij is the
distance between nodes i and j . We assume that data transmission is successful
only if the received power spectral density at the receiver exceeds a threshold
T . Likewise, we assume that an interference will become nonnegligible only
if it produces a power spectral density over a threshold of I at a receiver.
Based on the threshold T , the transmission range for a node is thus given
by RT = (/T )1/ , which is derived from (RT ) = T . Similarly,
based on the interference threshold I (< T ), the interference range for a node
is given by RI = (/I )1/ . Because I < T , we have RI > RT .
Scheduling and interference constraints Scheduling can be done either in
time domain or frequency domain. In this chapter, we consider a frequency
domain subband assignment, i.e., how to assign subbands at a node for trans-
mission and reception. A feasible scheduling on frequency bands must ensure
that there is no interference at the same node and among the neighboring
nodes.
Suppose that band m is available at both node i and node j , i.e., m

( (
Mi Mj . To simplify the notation, let Mij = Mi Mj . Denote
'
(m,k) 1 if node i transmits data to node j on subband (m, k),
xij =
0 otherwise.
For a node i N and a band m Mi , denote Tim as the set of nodes that
can use band m and are within the transmission range to node i, i.e.,
Tim = {j : dij RT , j = i, m Mj }.
We assume that node i cannot transmit to multiple nodes on the same frequency
sub-band. We have
(m,k)
xiq 1. (10.2)
qTim
For a frequency subband (m, k), if node i uses this subband for transmitting
data to a node j Tim , then any other node that can produce interference on
node j should not use this subband. Note that the so-called hidden terminal
problem is a special case under this constraint. To model this constraint, we
denote Ijm as the set of nodes that can produce interference at node j on band
m, i.e.,
Ijm = {p : dpj RI , Tpm = }.
The physical meaning of Tpm = in the above definition is that node p may
use band m for a valid transmission to some node in Tpm , which then causes
interference to node j . Thus, we have

xij(m,k) + (m,k)
xpq 1(p Ijm , p = i). (10.3)
qTpm
In (10.3), if xij(m,k) = 1, i.e., node i uses frequency subband (m, k) to trans-

mit to node j , then any node p that can cause interference on node j should
(m,k)
not transmit on this subband, i.e., qTpm xpq = 0. On the other hand, if
(m,k)
xij = 0, (10.3) degenerates into (10.2), i.e., node p may transmit on sub-
(m,k)
band (m, k) to at most one node q Tpm , i.e., qTpm xpq 1.
It is important to understand that in the interference constraint (10.3), if
(m,k)
xij = 0, two nodes that can produce interference at node j but are far apart
and outside each others interference range can use the same subband (m, k) for
transmission. We use an example to illustrate this point. In Fig. 10.2, suppose
that node 1 transmits to node 2 on subband (m, k). Then any node that can
cause interference at node 2 (i.e., node 3 or 5) cannot use the same subband
for transmission. On the other hand, if node 1 does not use subband (m, k)
to transmit to node 2, then node 3 may use this subband to transmit (to node
4) as stated in (10.3). Likewise, node 5 may also use this subband to transmit
Figure 10.2 An example

illustrating interference among
the links.
RI
6
5 2
RI
1 3
(to node 6) as stated in (10.3). That is, both nodes 3 and 5 may use the same
subband for transmission.
Flow routing At the network level, a source node may need a number of
relay nodes to route the data stream toward its destination node. Clearly, a route
having only a single path may be overly restrictive and may not be able to take
advantage of load balancing. A set of paths (with flow splitting) is more flexible
to route the traffic from a source node to its destination. Mathematically, this
can be modeled as follows. Denote fij (l) as the data rate on link i j that is

attributed to session l, where i N , j mMi Tim , and l L. To simplify

the notation, let Ti = mMi Tim . If node i is the source node of session l,
i.e., i = s(l), then

fij (l) = r(l). (10.4)
j Ti
If node i is an intermediate relay node for session l, i.e., i = s(l) and i = d(l),
then

fij (l) = fpi (l). (10.5)
j Ti ,j =s(l) pTi ,p=d(l)

fpi (l) = r(l). (10.6)
pTi
It can be easily verified that if (10.4) and (10.5) are satisfied, then (10.6) must
be satisfied. As a result, it is sufficient to list only (10.4) and (10.5) in the
formulation.
Link capacity constraint In addition to the above flow balance equations
at each node i for each session l, the aggregate flow rates on each radio link
cannot exceed this links capacity. To model this mathematically, we need to
first find the capacity on link i j and subband (m, k). If node i sends data
to node j on subband (m, k), i.e., xij(m,k) = 1, then the capacity on link i j
and subband (m, k) is given by
% &
(m,k) gij
cij = u(m,k) W (m) log2 1 + ,

where is the ambient Gaussian noise density. Note that the denominator
inside the log function contains only . This is due to one of our interference
constraints stated earlier, i.e., when node i transmits to node j on subband
(m, k), then all the other neighbors of node j within its interference range are
prohibited from using this sub-band. This interference constraint significantly
(m,k) (m,k)
helps simplify the calculation of the link capacity cij . When xij = 0, we
(m,k) (m,k)
have cij = 0. Thus, cij can be written in the following compact form:
% &
(m,k) (m,k) gij
cij = xij u (m,k)
W (m)
log2 1 + . (10.7)

Now, returning to our earlier requirement that the aggregate data rates on
each link i j cannot exceed the links capacity, we have
(m)
K (m,k)
fij (l) cij
lL,s(l)=j,d(l)=i mMij k=1
K
(m) % &
gij
= xij(m,k) u(m,k) W (m) log2 1 + .

mMij k=1
10.3.2 Problem formulation

For a multi-hop CRN of the type that we are investigating, various performance
objectives can be used. In this chapter, we use the same objective function as
in Chapter 5, i.e., the bandwidth-footprint-product (BFP). Since the footprint
is identical for all nodes, it can be removed without any loss of generality. It
is not hard to see that the solution procedure in this chapter can be likewise
applied for several other performance objectives as well.
Mathematically, we have the following optimization problem:
(m)
K
Minimize W (m) xij(m,k) u(m,k)
iN mMi j Tim k=1
(m)

K
subject to u(m,k) = 1 (m M)
k=1
(m,k)
xiq 1 (i N , m Mi , 1 k K (m) ) (10.8)
qTi m
253 10.4 Deriving a lower bound

xij(m,k) + (m,k)
xpq 1(i N , m Mi , j Tim ,
qTpm
1 k K (m) , p Ijm , p = i) (10.9)

lL K (m) % &
gij (m,k)
fij (l) W (m) log2 1+ xij u(m,k) 0

s(l)=j,d(l)=i mMij k=1
(i N , j Ti )

fij (l) = r(l) (l L, i = s(l))
j Ti
j =
s(l) p=
d(l)
fij (l) fpi (l) =0 (l L, i N , i = s(l), d(l))
j Ti pTi
(m,k)
xij = 0 or 1, u(m,k) 0 (i N , m Mi , j Tim , 1 k K (m) )
fij (l) 0 (l L, i N , i = d(l), j Ti , j = s(l)),
(m,k)
where W (m) , gij , , , and r(l) are all constants, and xij , u(m,k) , and fij (l)
are the optimization decision variables.
The above optimization problem is a mixed-integer nonlinear programming
(MINLP) problem, which is NP-hard in general [46]. Our approach to solve
this problem is as follows. In Section 10.4, we first propose the computation of
a lower bound for the problem, which can be obtained by relaxing the integer
variables and using a linearization technique. Using this lower bound as a per-
formance benchmark, we develop in Section 10.5 a highly effective algorithm
based on the SF procedure that we discussed in Section 10.1. Using exten-
sive simulation results, we show that the SF algorithm can offer solutions with
objective values very close to the computed lower bounds. Since the optimal
objective value lies between the lower bound and the solution obtained by the
SF algorithm, the solution produced by the SF algorithm must be even closer
to the optimum.
10.4 Deriving a lower bound
The complexity of the problem formulated in Section 10.3.2 arises from the
(m,k) (m,k)
binary xij -variables and the product of variables xij u(m,k) . To derive
a lower bound for the problem, we first multiply (10.8) and (10.9) by the
(m,k)
corresponding u(m,k) , so that xij appears throughout as a product with
(m,k)
u(m,k) . We then relax the integer (binary) requirement on xij with 0
xij(m,k) 1 and replace xij(m,k) u(m,k) with a single variable, say sij(m,k) , i.e.,
(m,k)
sij = xij(m,k) u(m,k) u(m,k) . Such a relaxation leads to the following lower-
bounding problem formulation:
(m)
K
Mininize W (m) sij(m,k)
iN mMi j Tim k=1
(m)

K
subject to u(m,k) = 1 (m M)
k=1
(m,k)
siq u(m,k) 0 (i N , m Mi , 1 k K (m) ) (10.10)
qTi m
(m,k)

sij + (m,k)
spq u(m,k) 0 i N , m Mi , j Tim ,
qTpm

1 k K (m) , p Ijm , p = i (10.11)

lL K
(m) % &
gij (m,k)
fij (l) W (m)
log2 1+ sij 0

s(l)=j,d(l)=i mMij k=1
(i N , j Ti )

fij (l) = r(l) (l L, i = s(l))
j Ti
j =
s(l) p=
d(l)
fij (l) fpi (l) = 0 (l L, i N , i = s(l), d(l))
j Ti pTi
(m,k)
u(m,k) , sij 0 (i N , m Mi , j Tim , 1 k K (m) )
fij (l) 0 (l L, i N , i = d(l), j Ti , j = s(l)).
This new (relaxed) formulation is a standard LP problem, the solution of which

can be obtained in polynomial time. Due to the relaxation (and thus enlarged
optimization space), the solution value to this LP problem yields a lower bound
for the original problem of Section 10.3.2.
10.5 A near-optimal algorithm based on sequential fixing

10.5.1 Basic algorithm
We now take a closer look at the original MINLP problem formulation in Sec-
tion 10.3.2. Observe that once the binary values for all the x-variables are
determined, i.e., we have ascertained whether or not a node will indeed use
a particular subband to send data to another node, then this MINLP reduces
to an LP, which can be solved in polynomial time. Thus, the key obstacle in
solving this MINLP problem lies in the determination of the binary values for
all the x-variables. To this end, we propose the following two-step solution
procedure:
255 10.5 A near-optimal algorithm based on sequential fixing
1. Fix the binary values for the x-variables iteratively through a sequence of
LPs.
2. Once all the x-variables are fixed, find a solution (to determine how to
divide sub-bands and the flow routing) based on the x-variable values
obtained in Step 1.
Such a two-step approach will yield a suboptimal (upper bounding) solution to

the original MINLP problem. The quality of this algorithm can be assessed by
comparing the resulting objective value to the lower bound that we derived in
the previous section.
As said, the key to the two-step approach resides in the determination of the
(m,k)
binary values for all the xij -variables. Our main idea is to fix the values
of the xij(m,k) -variables sequentially by solving a series of relaxed LP prob-
lems and setting at least one binary value for some xij(m,k) -variable during
each iteration. Specifically, during the first iteration, we relax all the binary
variables xij(m,k) to satisfy 0 xij(m,k) 1, as in Section 10.4, to derive an
LP problem. Upon solving this LP problem, we obtain a solution with each
(m,k)
xij = sij(m,k) /u(m,k) being a value between zero and one. Among all these
x-variable values, we select some xij(m,k) that has the largest value. Then we
(m,k)
fix (set) this particular xij to 1. As a result of this fixing, by (10.8), we also
(m,k)
need to fix xiq = 0 for q Tim and q = j . Further, by (10.9), we can fix
(m,k)
xpq to 0 for p Ijm , p = i, and q Tpm .
Algorithm 10.1 An SF algorithm

1. Set up and solve the initial relaxed LP problem, as shown in Section 10.4.
2. Suppose xij(m,k) has the largest value among all the x-variables that
(m,k)
remain to be fixed; fix this xij = 1.
(m,k) (m,k)
Also, fix xiq = 0 (for q Ti and q = j ) and xpq
m = 0 (for p
Ij , p = i, and q Tp ).
m m
(m,k)
3. If all the xij -variables are fixed, go to Step 5.
4. Reformulate and solve a new relaxed LP problem with the newly fixed
x-variables and go to Step 2.
(m,k)
5. Formulate and solve the LP problem based on the fixed xij -variable
values.
Now, having fixed some x-variables in the first iteration, we update the prob-
lem to obtain a new LP problem for the second iteration as follows. For those
(m,k) (m,k)
xij -variables that are already fixed to one, since sij = xij(m,k) u(m,k) =
u(m,k) , we can replace the corresponding sij(m,k) by u(m,k) . For those xiq
(m,k)
and
(m,k) (m,k) (m,k)
xpq that are fixed to zero, we can set siq = 0 and spq = 0. As a result,
Figure 10.3 100 data sets of 2

normalized objective (with
respect to the computed lower 1.8
bound) for 20-node networks. 1.6
1.4
Normalized objective
(w.r.t. lower bound)
1.2
0.8
0.6
0.4
0.2
0
1 10 20 30 40 50 60 70 80 90 100
Set index
all the terms in the LP problem involving these s-variables can be removed and
the corresponding constraints in (10.10) and (10.11) can also be removed.
In the second iteration, we solve this new LP problem and then fix some
additional x-variables based on the same process (now the ordering of the
x-values is done only for the remaining not-yet-fixed or free x-variables).
The iteration continues and eventually we fix all the x-variables to either zero
and one.
Upon fixing all the x-variable values, the original MINLP reduces to an LP
problem, which can be solved in polynomial time. The complete SF algorithm
is displayed in Fig. 10.3.
10.5.2 A speedup technique

In the SF algorithm, we need to solve a polynomial sequence of LPs, which
yields polynomial-time complexity. By exploiting the spatial and spectral
dimensions involved in radio resource allocation, we may decrease the number
of LPs by fixing more x-variables during each iteration in Fig. 10.3. As a result,
the number of iterations can be further decreased. From the spatial dimension,
a subband usage will only have an impact within the interference range, and the
same subband can be used by other links outside this range. Thus, for the same
subband (m, k), we may fix multiple links that have nonoverlapping interfer-
ence ranges within a single iteration of the SF algorithm. From the spectral
dimension, the transmission on one subband will not interfere with the trans-
mission on a different subband. Thus, for the same link i j , we may fix
multiple subbands within a single iteration of the SF algorithm. Specifically,
we can use a threshold > 0.5 in this fixing process and fix all the x-variables
that exceed to one in a single iteration. Note that in (10.8) and (10.9), it is
Table 10.2 Available bands M in the network in the simu-

lation study.
Band index Spectrum range (MHz) Bandwidth (MHz)
I [1240, 1300] 60.0

II [1525, 1710] 185.0
III [902, 928] 26.0
IV [2400, 2483.5] 83.5
V [5725, 5850] 125.0
required that at most one binary x-variable equals one, whereas in the relaxed
(m,k)
problem, there is at most one fraction siq /u(m,k) that exceeds 0.5. Thus,
> 0.5 ensures that both the constraints (10.8) and (10.9) (interference con-
straints at each node and among the nodes) will hold during the SF procedure.
We used = 0.85 in our numerical results in the next section. In the case that
none of the x-variables exceeds , we will fall back to the basic algorithm in
Fig. 10.3 and simply choose the largest valued x-variable.
In this section, we present numerical results for our SF algorithm and com-
pare it to the lower bound that we obtained in Section 10.4. We consider
|N | = 20, 30, or 40 nodes in a 500 500 area (in meters). Among these nodes,
there are |L| = 5 active sessions, each having a rate that is randomly generated
within [10, 100] Mb/s.
We assume that there are M = 5 bands that can be used for the entire net-
work (see Table 10.2). Bands I and II are among the least-utilized spectrum
bands found in [106] (less than 2%), and bands III, IV, and V are unlicensed
ISM bands used for 802.11. Recall that the set of available bands at each CR
node is a subset of these five bands based on the nodes location, and that the
set of available bands at any two nodes in the network may not be identical. In
the simulation, this is done by randomly selecting a subset of bands from the
pool of five bands for each node. Further, we assume that bands I to V can be
divided into 3, 5, 2, 4, and 4 subbands, although other desirable divisions can
be used. Note that the size of each subband may be unequal and is part of the
optimization problem.
We assume that the transmission range at each node is 100 m and that the
interference range is 150 m, although other settings can be used. The path-loss
index is assumed to be 4 andwe set = 62.5. The threshold T is assumed
to be 10. Thus, we have I = 150 T , and the transmission power spectral
100
density is given by = (100) T / = 1.6 107 .

Note that it is possible that there is no feasible solution for a specific data
set. This could be attributed to the loss of connectivity in certain areas in the
network (due to random network topology), resource bottleneck in a hot area,
etc. Thus, we only report results based on those data sets that have feasible
solutions.
We first present simulation results for 100 data sets for 20-node networks
that can produce feasible solutions. For each data set, the network topology,
the sourcedestination pairs and bit rates of all sessions, and the available fre-
quency bands at each node are randomly generated. We used the SF algorithm
to determine a feasible solution, and we compared its objective value with the
lower bound derived as in Section 10.4. The run time consumed by each sim-
ulation was less than 10 seconds on a Pentium 3.4 GHz machine.
Fig. 10.3 depicts the normalized objective values obtained by the SF algo-
rithm with respect to the computed lower bounds for 100 data sets. The average
normalized objective value obtained among the 100 simulations is 1.04 and the
standard derivation is 0.07. There are two observations that can be made from
this figure. First, since the ratio of the solution value obtained by SF (upper
bound on the optimal solution value) to the lower bound value is close to 1
(in many cases, they coincide with each other), the lower bound must be very
tight. Second, since the actual (unknown) optimal solution value lies between
the solution value obtained by the SF algorithm and the lower bound, the SF
solution value must be even closer to the optimal value than that indicated by
the foregoing ratio.
To get a sense of how the actual (rather than the normalized) numerical
results appear in the simulations, we list the 100 sets of results in Table 10.3.
Note that in many cases, the result obtained by the SF algorithm is identical
to the respective lower bound obtained via the underlying relaxation, which
indicates that the solution found by SF in such cases is optimal.
Simulation results for the 100 random data sets corresponding to the 30-
node and the 40-node networks that produce feasible solutions are shown in
Figs. 10.4 and 10.5, respectively. For the 30-node networks, the average nor-
malized value obtained was 1.10, with a standard derivation of 0.16. For the
40-node networks, the average normalized value obtained was 1.18, with a
standard derivation of 0.16. Thus, the derived SF solutions are again close to
optimality for these data sets.
In this chapter, we presented an effective approach to address a class of

mixed-integer optimization problems. The technique, called sequential fix-
ing is designed to iteratively determine (fix) binary integer variables. It is a
heuristic procedure and has a polynomial-time complexity. Its performance is
Table 10.3 Simulation results (in MHz) of 100 data sets for 20-node networks.
Set Lower Result Set Lower Result Set Lower Result

index bound by SF index bound by SF index bound by SF
1 138.33 138.33 35 270.76 319.71 68 214.82 214.82

2 156.12 156.12 36 325.59 394.43 69 177.85 177.85
3 173.53 173.53 37 288.72 288.72 70 233.50 264.12
4 189.70 189.70 38 244.77 247.74 71 298.32 298.32
5 203.05 213.18 39 215.72 223.83 72 215.45 215.45
6 184.37 184.37 40 126.05 126.05 73 246.76 248.23
7 160.45 182.33 41 191.79 191.79 74 270.04 270.04
8 232.23 232.23 42 171.11 171.11 75 163.74 165.16
9 223.00 223.53 43 157.95 157.95 76 213.42 213.42
10 182.13 182.13 44 238.02 238.02 77 151.64 151.64
11 220.20 220.20 45 129.96 219.96 78 369.62 369.62
12 277.83 277.83 46 244.70 254.07 79 110.70 114.29
13 130.54 134.05 47 293.82 293.82 80 180.11 180.11
14 172.62 172.62 48 372.99 396.06 81 162.15 162.15
15 256.96 256.96 49 240.53 301.21 82 183.67 192.64
16 178.73 178.73 50 367.20 472.34 83 366.16 430.61
17 152.08 152.08 51 128.39 128.39 84 130.76 130.76
18 359.03 359.03 52 238.61 238.61 85 420.90 421.74
19 150.61 150.61 53 127.82 127.82 86 177.20 234.00
20 164.97 164.97 54 195.19 205.00 87 205.77 205.77
21 156.43 156.43 55 208.53 232.90 88 211.40 211.40
22 238.41 308.51 56 188.86 189.51 89 226.85 242.22
23 184.78 184.78 57 170.53 188.55 90 164.47 175.27
24 241.42 243.22 58 183.51 195.68 91 203.80 203.80
25 135.39 140.96 59 376.52 481.59 92 267.63 267.63
26 247.30 251.18 60 265.62 265.62 93 157.91 167.85
27 280.80 290.85 61 209.20 209.20 94 303.65 308.71
28 353.98 354.17 62 139.66 139.73 95 225.93 230.63
29 260.56 260.56 63 168.52 184.07 96 194.93 224.04
30 127.06 127.06 64 111.29 111.29 97 160.25 160.25
31 170.35 170.35 65 118.07 118.07 98 165.02 165.02
32 207.74 207.74 66 179.47 186.34 99 99.79 99.79
33 183.59 183.59 67 100.05 100.05 100 168.68 168.68
34 138.33 143.00
typically measured by comparing its solution value to some performance

bound, e.g., a lower bound for a minimization problem, or an upper bound
for a maximization problem. Based on our own experience, we found that that
this SF technique is very efficient and can offer highly competitive solutions.
As a case study, we studied an optimization problem in a multi-hop CRN.
Since the problem formulation is an MINLP, we developed a lower bound
to estimate the optimal objective value. Subsequently, we presented an SF
algorithm for this optimization problem. Numerical examples showed that the

1.4
(w.r.t. Lower bound)
1.2
0.8
0.6
0.4
0.2
0
1 10 20 30 40 50 60 70 80 90 100
Set index

1.4
(w.r.t. lower bound)
1.2
0.8
0.6
0.4
0.2
0
1 10 20 30 40 50 60 70 80 90 100
Set index
solutions produced by this SF algorithm can offer objective values that are very
close to the computed lower bounds, thus confirming their near-optimality.
10.8 Problems
10.1 For the problem considered in this chapter, what is the purpose of
developing a lower bound? If a solution obtained by the SF algorithm is very
close to the lower bound, what can we claim?
261 10.8 Problems
10.2 Why is it necessary to divide a large band into smaller subbands in a

CRN?
10.3 Show that, for a session, if flow balance holds at the source node
(Eq. (10.4)) and at the relay nodes (Eq. (10.5)), then flow balance also holds at
the sessions destination node (Eq. (10.6)).
10.4 In this chapter, we used the so-called network-wide bandwidth-
footprint-product (BFP) as the objective function. However, it is possible that
two or more neighboring nodes that transmit on the same frequency band have
some partial overlap of their footprints. As a result, the overlapped area will
be counted multiple times in the objective function. Provide a justification on
why this objective function still makes sense.
(m,k)
10.5 The capacity cij on link i j and subband (m, k) can be expressed
as
% &
(m,k) gij
cij =u (m,k) m
W log2 1+ .

Referring to the RHS, explain why there is no interference term in the denom-
inator inside the log function when calculating the SINR.
10.6 For the formulation of the original problem presented in this chapter,
classify the constraints based on the different layers. Which constraints couple
multiple layers?
(m,k)
10.7 In the SF algorithm, we fix the xij -variable having the largest value
among all undetermined x-variables to 1 during each iteration. Comment on
(m,k)
the alternative strategy of fixing the xij -variable having the smallest value
among all undetermined x-variables to 0. Which is a more effective technique
and why?
10.8 To speed up variable fixing, a threshold is used to fix several undeter-
mined x-variables to 1. Discuss the pros and cons of setting a large vs a small
value of .
10.9 In the numerical examples, the solution obtained by the SF algorithm
coincides with the lower bound solution in many cases. What can we infer from
this observation and why? Also, if the SF algorithm fails to find a solution, can
we claim that the original problem is infeasible? Why, or why not?
10.10 Explain why the SF algorithm is of polynomial-time complexity.
Derive an order of complexity for this algorithm.
10.11 Consider a mixed-integer optimization problem where the goal is to
maximize an objective function rather than to minimize it. Describe how to
design an SF algorithm for this maximization problem.
CHAPTER
11 Metaheuristic methods
One day your life will flash before your eyes. Make sure it is worth
watching.
Unknown
11.1 Review of key results in metaheuristic methods
In this chapter, we discuss another class of heuristics, which are known as

metaheuristic methods [36]. An iteration in metaheuristic methods typically
aims to improve the current feasible solution, with the initial solution given by
the user. Some well-known metaheuristic methods are iterative improvement,
simulated annealing, tabu search, and genetic algorithms [36]. For certain type
of problems, metaheuristic methods could be very effective.
The so-called iterative improvement (or basic local search) method tries to find
a better solution in each iteration by searching in the neighborhood of the current
solution, and terminates when a better solution cannot be found. It has been
shown that the performance of iterative improvement methods for combinatorial
optimization problems may not be satisfactory [19]. This can be explained by
the fact that this method tends to stop as soon as it finds a local optimum.
Compared to iterative improvement, simulated annealing (SA) [1] has an
explicit strategy to escape from local optima. The basic idea of SA is to allow a
move (with a probability) even if it may tentatively result in a solution of worse
quality than the current solution. There is also a cooling procedure in SA,
which decreases such randomness (or diversification) as time passes. As the
cooling proceeds, SA gradually converges to a simple iterative improvement
algorithm, which guarantees convergence. The performance of SA is sensitive
to the initial solution and the neighborhood structure (in addition to the cooling
procedure).
262
263 11.1 Review of key results in metaheuristic methods
Figure 11.1 Flow chart of GA.

Representation
and
Initialization
Evaluation
Meet Yes
termination
criteria
No
Selection Stop
q (1q)
Crossover
m
(1m)
Mutation
Another well-known technique is tabu search (TS), which explicitly uses the
history of the search (recorded in a list of solutions declared as tabu) both
to escape from local optima and to implement an exploratory strategy [51].
Similar to SA and iterative improvement, the TS approach is a single-point-
search metaheuristic (also called trajectory methods). The performance of
TS strongly depends on the initial solution and the neighborhood structure,
and on the diversification and intensification schemes that it uses to explore
new or previous promising portions of the search region, respectively.
Unlike the above three metaheuristic methods, a genetic algorithm (GA) is a
population-based algorithm that is inspired by the survival-of-the-fittest prin-
ciple , as derived from its natural evolution context. Because of this algorithmic
structure, a GA is also classified as an evolutionary algorithm. This procedure
has the intrinsic strength of dealing with a set of solutions (i.e., a population) at
each generation rather than working with a single, current solution. To repre-
sent a solution in a GA, we need to first define a data structure (i.e., a gene) and
then map a solution to a sequence of genes (i.e., a chromosome). A GA needs a
set of solutions (or individuals) as the first generation elements. This is the ini-
tial step in Fig. 11.1. A GA evaluates each individual by a fitness function (in
the form of an objective function). If the set of current individuals meets some
termination criteria, then the procedure stops and returns the best individual
as the final solution. Otherwise, a GA applies a number of genetic operators

to the current individuals such that better individuals can be generated for the
next generation. The fitness values of the first generation solutions strongly
influence the complexity of GA. That is, if we can identify a set of individuals
having good fitness values for the first generation, then we may need a fewer
number of iterations to derive a final solution.
In the next iteration, GA selects good individuals as parents based on their
fitness values. The basic assumption here is that a good solution often con-
tains some good genes that may also be shared by the optimal solution. To
remove bad genes from the population, individuals with low fitness values will
be removed out of the population. Then a GA uses a genetic operator known as
crossover (with a probability ) to recombine two or more selected individuals
to produce new individuals that share good genes from the selected individuals.
A GA also uses a mutation operator (with a probability ) to achieve a random-
ized self-adaptation of individuals. Each new individual is evaluated and may
be placed in the next population group if it has a good fitness value. Based
on the survival-of-the-fittest principle, the overall quality of the population is
likely to improve as the algorithm progresses from one generation to the next.
In this chapter, we will illustrate how a metaheuristic method such as GA
can be applied to solve some complex combinatorial optimization problems in
wireless networks.
11.2 Case study: Routing for multiple description video over

wireless ad hoc networks
As progress in wireless ad hoc networking continues, there is an increasing
expectation on enabling content-rich video communications in such networks.
However, there are significant technical barriers that hinder the deployment of
video applications in wireless ad hoc networks. In fact, what makes traditional
single stream coding and layered coding successful in the internet and certain
wireless networks is the existence of a relatively stable path during the video
session. Consequently, packet loss on important information (e.g., base layer)
is kept low, and can be effectively controlled by error control and concealment
mechanisms. This is important since, for a layered video, the successful recon-
struction of a video relies on the base layer, and the decoding of enhancement
layers hinges upon lower enhancement layers as well as upon the base layer.
However, this situation hardly holds true in wireless ad hoc networks, where
there may not exist any single reliable path, and packet loss may be beyond the
recovery capability of most error control mechanisms.
Recently, multiple description (MD) coding has become a popular coding
technique for media streaming [57]. With MD coding, multiple streams (or
descriptions) are generated for a video source and any received subset of these
streams can be used to reconstruct the original video. Further, the quality of
received video commensurate with the number of received descriptions. Such
flexibility has positioned MD coding as a perfect candidate for video commu-
nications over wireless ad hoc networks [102]. This is because the topology of
such networks is intrinsically mesh, within which multiple paths exist between
any source and destination pair. Although the links in such networks are fragile,
the probability of concurrent loss of all of the descriptions is low. Therefore,
MD coding is likely to be very effective.
In this chapter, we study the problem of multi-path routing for double
description (DD) video in wireless ad hoc networks. We follow a cross-layer
approach in problem formulation by considering the application layer perfor-
mance (i.e., average video distortion) as a function of lower layer parameters
(e.g., bandwidth, loss, and path correlation). We show that the objective func-
tion is a complex ratio of high-order exponentials that is not decomposable.
Consequently, it is futile to develop a tractable analytic solution.
However, we find that a metaheuristic technique such as a GA is eminently
suitable in addressing such type of complex cross-layer optimization problems.
This is because GAs possess an intrinsic capability of handling a population of
solutions (rather than working with a single current solution during each itera-
tion). Such capability gives GAs the unique strength of identifying promising
regions in the search space (not necessarily convex) and having less of a ten-
dency to be trapped in a local optimum, as compared with other trajectory-
based metaheuristics (e.g., simulated annealing (SA) and tabu search (TS)
[19]). Using numerical examples, we show that significant performance gains
can be achieved by a GA-based approach over trajectory-based approaches.
formulate a cross-layer optimization problem for DD video over multiple paths
in ad hoc networks. In Section 11.4, we describe our GA-based solution. In
Section 11.5, we present numerical examples. Section 11.6 summarizes this
chapter.
11.3 Problem description
An ad hoc network can be modeled as a stochastic directed graph G{V , E},

where V is the set of vertices and E the set of edges. We assume that nodes are
reliable during the video session, but links may fail with certain probabilities.
Further, we assume that orthogonal channels are used among the links so that
interference among the links can be effectively managed by proper channel
assignment. We characterize a link {i, j } E by:
bij : the available bandwidth of link {i, j }.

pij : the probability that link {i, j } is up.
lij : average burst length in a packet loss on link {i, j }.
Table 11.1 Notation
Symbol Definition
bij Bandwidth of link {i, j }

Bj nt Minimum bandwidth of the shared links
d0 Distortion when both descriptions are received
dh Distortion when only description h is received, h = 1, 2
D Average distortion
E Set of edges in the network
gi An intermediate node in a path
G{V , E} Graph representation of the network
{i, j } A link from node i to node j
lij Average length of loss burst on link {i, j }
pij Success probability of link {i, j }
pj nt Average success probability of joint links
h
pdj Average success probability of disjoint links on Ph
P A path from s to t
P00 Probability of receiving both descriptions
P01 Probability of receiving description 1 only
P10 Probability of receiving description 2 only
P11 Probability of losing both descriptions
Rh Rate of description h in bits/sample
s Source node
t Destination node
Ton Average up period of the joint links
V Set of vertices in the network
h
xij Routing index variables, defined in (11.10)
ij up to down transition probability of link {i, j }
ij down to up transition probability of link {i, j }
GA crossover rate
GA mutation rate
From these link-level statistics, we can explore path-level bandwidth and fail-
ure probabilities, which are the key factors to determine video distortion (see
(11.2)). Other link characteristics, such as delay, jitter, congestion, and signal
strength may also be incorporated into this framework as well (e.g., see [103]).
Table 11.1 lists the notation used in this chapter.
11.3.1 Rate-distortion regions for DD coding

Throughout this chapter, we use DD coding for MD video, which is widely
used in practice [8; 12; 23; 53; 102]. In general, using more descriptions
and paths will increase the robustness to packet loss and path failure. But
more descriptions may increase the video bit rate for the same video qual-
ity. A study in [135] demonstrates that the most significant performance gain
is achieved when the number of descriptions increases from one to two, but
with only marginal improvements achieved for further increase in the number
of descriptions.
For video coding and communications, a distortion rate model describes the
relationship between the achieved distortion and the bit rate (i.e., the quality
and the length of the representation). For two descriptions (each generated for
a sequence of video frames), denote dh as the achieved distortion when only
description h is received, h = 1, 2, and d0 the distortion when both descrip-
tions are received. Denote Rh as the rate in bits/pixel of description h, h = 1, 2.
The rate-distortion region for a memoryless independent and identically dis-
tributed (i.i.d.) Gaussian source with the square error distortion measure was
first introduced in [117]. For computational efficiency, Alasti et al. in [6] intro-
duce the following distortion-rate function, which we use in this chapter:

22(R1 +R2 )
d0 = 22R1 +22R2 22(R1 +R2 )
2
d1 = 22R 1 2 (11.1)

2R
d2 = 2 2 2,
where 2 is the variance of the source. Denote P00 as the probability of receiv-
ing both descriptions, P01 as the probability of receiving description 1 only,
P10 as the probability of receiving description 2 only, and P11 as the proba-
bility of losing both descriptions. Then, the average distortion of the received
video can be approximated as:
D = P00 d0 + P01 d1 + P10 d2 + P11 2 . (11.2)
Our simulation results show that although the distortion-rate function in

(11.1) is an approximation for DD video, significant improvement in received
video quality could be achieved over alternative approaches by incorporating
(11.2) into the optimal routing problem formulation (see Section 11.5). Fur-
ther, our formulation does not depend on any specific distortion-rate function.
A more accurate distortion-rate function for MD video could be incorporated
into our formulation should it become available in the future.
11.3.2 Description rates and success probabilities

As a first step to formulate the problem of optimal multi-path routing, we need
to know how to compute the average distortion D as a function of link statistics
for a given pair of paths. That is, we need to compute the end-to-end bandwidth
(or rate) for each stream and joint probabilities of receiving the descriptions
(see (11.1) and (11.2)).
For a sourcedestination pair {s, t}, consider two given paths [P1 , P2 ] in
G{V , E}. Since we do not mandate disjointedness in routing, P1 and P2 may
share nodes and links. Similar to the approach in [8] and [12], we classify the
links along the two paths into three sets: set one consisting of links shared by
both paths, denoted as J (P1 , P2 ), and the other two sets consisting of disjoint
links on the two paths, denoted as J (Ph ), h = 1, 2, respectively. Then, the

minimum bandwidth of J (P1 , P2 ), Bj nt , is
'
min{i, j } J (P1 , P2 ) {bij }, if J (P1 , P2 ) = ,
Bj nt = (11.3)
otherwise.
The rates of the two video streams, R1 and R2 , can be computed as

'
Rh = B(Ph ), if 2m=1 B(Pm ) Bj nt , h = 1, 2,
(11.4)
R1 + R2 Bj nt otherwise,
where B(Ph ) = min{i,j }Ph {bij }, h = 1, 2, and is a constant determined by

the video format and frame rate. For a video with coding rate f frames/s and
a resolution of W V pixels/frame, we have = 1/( W V f ), where
is a constant determined by the chroma subsampling scheme. For the quar-
ter common intermediate format (QCIF) [176 144 Y pixels/frame, 88 72
Cb/Cr pixels/frame], we have = 1.5 and = 1/(1.5 176 144 f ). The
first line in (11.4) is for the case when the joint links are not the bottleneck
of the paths. The second line of (11.4) is for the case where one of the joint
links is the bottleneck of both paths. In the latter case, we assign the bandwidth
to the paths by splitting the bandwidth of the shared bottleneck link in propor-
tion to the mean success probabilities of the two paths, while an alternative
approach is to split the bandwidth evenly for balanced descriptions.
We now focus on how to compute the end-to-end success probabilities. For
disjoint portion of the paths, it suffices to model packet loss as a Bernoulli
event. This is because the loss of the two descriptions is assumed to be inde-
pendent in the disjoint portions. Thus, the success probability on the disjoint
portions of the two paths is
'
{i, j } J (Ph ) pij , if J (Ph ) = , h = 1, 2,
h
pdj = (11.5)
1 otherwise, h = 1, 2.
On the joint portion of the paths, loss of the two streams is correlated.
To model such correlation, we assume that each shared link {i, j } follows
an onoff process modulated by a discrete-time Markov chain, as shown in
Fig. 11.2(a). With this model, there is no packet loss when the link is in
the up state; all packets are dropped when the link is in the down state.
Transition probabilities, {ij , ij }, can be computed from the link statistics as
ij = 1/ lij and ij = (1 pij )/(pij lij ).
For K shared links, the aggregate failure process of these links is a Markov
process with 2K states. In order to simplify the computation, we follow the
well-known Fritchman model [43] to model the aggregate process as an onoff
process. Since a packet is successfully delivered on the joint portion if and only
if all joint links are in the up state, we can lump up all the states with at least
one link failure into a single down state, while using the remaining state
Figure 11.2 Link and path aij

models.
1aij UP DOWN 1bij
bij
(a) The Gilbert two-state link model.
J(P1)
(1)
1 p1dj

S UP DOWN t
(1b)
1 p2dj
J(P1,P2)
J(P2)
(b) A simplified path model for double-description video.
where all the links are in good condition as the up state. Denote Ton as the
average length of the up period. We have
1
Ton = . (11.6)
1 {i, j } J (P1 , P2 ) (1 ij )
If the joint link set is not empty, the probability of a successful delivery on the
joint links can be written as
'
{i, j } J (P1 , P2 ) pij , if J (P1 , P2 ) = ,
pj nt = (11.7)
1 otherwise.
Finally, the transition probabilities of the aggregate onoff process are
1 pj nt
= and = . (11.8)
Ton Ton (1 pj nt )
Note that = 0 and = 0 if J (P1 , P2 ) = . The consolidated path model is

illustrated in Fig. 11.2(b), where J (P1 , P2 ) is modeled as a two-state Markov
process with parameters {, }, and J (Ph ) is modeled as a Bernoulli process
with parameter (1 pdj h ), h = 1, 2.
With the consolidated path model, the joint probabilities of receiving the
descriptions are

P00 = pj nt (1 ) pdj
1 p2

! dj "

P01 = pj nt pdj
1 1 (1 ) p 2
! " dj
(11.9)

P10 = pj nt 1 (1 )pdj 1 p2

! dj "

P11 = 1 pj nt p 1 + p2 (1 ) p 1 p 2 .
dj dj dj dj
11.3.3 The optimal multi-path routing problem

With the above preliminaries, we now set out to formulate the multipath rout-
ing problem for MD video. To characterize any st path Ph , we define the
following binary variables:
'
1, if {i, j } Ph ,
xij =
h
(11.10)
0 otherwise.
With these variables, an arbitrary path Ph can be represented by a vector xh of

|E| elements, each of which corresponds to a link and has a binary value. We
formulate the problem as follows:
OPT-MM
Minimize D = P00 d0 + P01 d1 + P10 d2 + P11 2 (11.11)

1, if i = s, i V , h = 1, 2,
subject to xij
h
xj i = 1, if i = t, i V , h = 1, 2,
h

j :{i,j }E j :{j,i}E 0 otherwise, i V , h = 1, 2
(11.12)
'
1, if i = t, i V , h = 1, 2
xijh (11.13)
= 0, if i = t, i V , h = 1, 2
j :{i,j }E
xij1 R1 + xij2 R2 bij , {i, j } E (11.14)

xijh {0, 1}, {i, j } E, h = 1, 2. (11.15)
In Problem OPT-MM, {xijh } are binary optimization variables. Constraints

(11.12) and (11.13) ensure that the paths are loop-free, while constraint (11.14)
ensures the links are stable. For a given pair of paths, the average video dis-
tortion D is determined by the end-to-end statistics and the correlation of the
paths, as given in (11.1), (11.4), and (11.9). Different statistics of a given pair
of paths lead to a different video distortion. Specifically, the larger the end-to-
end bandwidth is, the higher the video rate is, and the smaller the distortion is.
With a lower end-to-end loss rate, fewer video frames will be corrupted. This
is modeled in (11.11), where 2 is usually much larger than d0 , d1 , and d2 , and
dh is usually larger than d0 , h = 1, 2. Finally, the impact of path correlation
is actually considered in the derivation of the joint probabilities of receiving
the description. In Problem OPT-MM, all these three elements are integrated
in the objective function (11.11), and are jointly optimized in routing.
The objective function (11.11) is a highly complex ratio of high-order
exponentials of the x-variables. The objective evaluation of a pair of paths
involves identifying the joint and disjoint portions, which is only possible when
both paths are completely determined (or can be conditioned on the exceed-
ingly complex products of the binary factors xij1 and (1 xij1 ) with xij2 and
(1 xij2 )). In [145], Sherali et al. considered a problem that seeks a pair of
disjoint paths in a network such that the total travel time over the paths is
271 11.4 A metaheuristic approach
minimized, where the travel time on a link might be either a constant or a

non-decreasing (or unstructured) function of the time spent on the previous
links traversed. Even for a simple special case where all the links except one
have a constant travel time (and hence linear objective terms), the problem
was shown to be NP-hard. Our problem has much more complex relationships
pertaining to the contribution of each individual link to the objective function,
which depends in general on the other links that are included in a fashion that
has no particular structural property such as convexity. Hence, it is likely to be
NP-hard as well. However, we do not have a proof in this chapter.
11.4 A metaheuristic approach
In this section, we present a solution procedure that produces a pair of feasible

and near-optimal paths.
We find that GAs [9] are eminently suitable for addressing this type of com-
plex combinatorial problems, most of which are multi-modal and nonconvex.
GAs are population-based metaheuristic inspired by the survival-of-the-fittest
principle. It has the intrinsic strength of dealing with a set of solutions (i.e.,
a population) at each step rather than working with a single, current solution.
At each iteration, a number of genetic operators are applied to the individuals
of the current population in order to generate individuals for the next genera-
tion. In particular, GA uses genetic operators known as crossover to recombine
two or more individuals to produce new individuals, and mutation to achieve
a randomized self-adaptation of individuals. The driving force in GA is the
selection of individuals based on their fitness (in the form of an objective func-
tion) for the next generation. The survival-of-the-fittest principle ensures that
the overall quality of the population improves as the algorithm progresses from
one generation to the next.
We follow the flow chart in Fig. 11.1 to design a GA-based solution to the
MD multi-path routing problem. In what follows, we use an example ad hoc
network shown in Fig. 11.3(a) to illustrate the components in a GA-based
solution. The termination condition in Fig. 11.1 could be based on the total
number of iterations (generations), maximum computing time, a threshold of
desired video distortion, or a threshold based on the lower bound developed
in [104].
11.4.1 Solution representation and initialization

In a GA, a feasible solution is encoded in the genetic format. For a routing
problem, a natural encoding scheme would be to define a node as a gene. Then,
an end-to-end path, consisting of an ordered sequence of nodes (connected by
the corresponding wireless links), can be represented as a chromosome [2].
Figure 11.3 Example network 2 6

and coding of an individual
(s = 1 and t = 9). 1 3 5 7 9
4 8
(a) Example ad hoc network.
An individual
P1 1 3 5 7 9
P2 1 2 5 7 8 9
(b) Example individual.
For Problem OPT-MM, each feasible solution consists of a pair of paths (i.e., a
pair of chromosomes), denoted as [P1 , P2 ]. An individual in this case is a pair
of vectors containing the nodes on paths P1 and P2 (see, e.g., Fig. 11.3(b)).
Before entering the main loop in Fig. 11.1, we need to generate an initial
population, i.e., a set of solutions. A simple approach would be to generate this
set of solutions by randomly appending feasible elements (i.e., nodes with con-
nectivity) to a partial solution. Under this approach, each construction process
starts with source node s. Then, the process randomly chooses a link incident
to the current end-node of the partial path and appends the link with its corre-
sponding head-node to augment the path, until destination node t is reached. It
is important to ensure that the intermediate partial path is loop-free during the
process. After generating a certain set of paths for st independently, a pop-
ulation of individuals can be constructed by pairing paths from this set. Our
numerical results show that a properly designed GA is not very sensitive to the
quality of the individuals in the initial population.
11.4.2 Evaluation
The fitness function f (x) of an individual, x = [P1 , P2 ], is closely tied to the
objective function (i.e., distortion D). Since the objective is to minimize the
average distortion function D, we use a fitness function which is defined as
the inverse of the distortion value, i.e., f (x) = 1/D(x). This simple fitness
definition appears to work very well, although other better fitness definitions
may be available.
11.4.3 Selection
During this operation, GA selects individuals that have a better chance or
potential to produce good offspring in terms of their fitness values. By
virtue of the selection operation, good genes among the population are more
likely to be passed to the future generations. We use the so-called Tournament
273 11.4 A metaheuristic approach
Figure 11.4 An example of the
Offspring 1
Parent 1
P1 1 3 5 7 9 P1 1 3 5 7 9
crossover operation.
P2 1 2 5 7 8 9 P23 1 2 5 8 9
Parent 2
Offspring 2
P3 1 4 5 8 9 P32 1 4 5 7 8 9
P4 1 3 4 5 6 9 P4 1 3 4 5 6 9
selection [9] scheme, which randomly chooses m individuals from the popula-
tion each time, and then selects the best of these m individuals in terms of their
fitness values. By repeating either procedure multiple times, a new population
can be selected.
11.4.4 Crossover
Crossover mimics the genetic mechanism of reproduction in the natural world,
in which genes from parents are recombined and passed to offspring. The deci-
sion of whether or not to perform a crossover operation is determined by the
crossover rate .
Fig. 11.4 illustrates one possible crossover implementation. Suppose that
we have two-parent individuals x1 = [P1 , P2 ] and x2 = [P3 , P4 ]. We could
randomly pick one path in x1 and one in x2 , say P2 and P3 . If one or more
common nodes exist in these two chosen paths, we could select the first such
common node that exists in P2 , say gr , where gr / {s, t}, and we can then
concatenate nodes {s, . . . , gr } from P2 with nodes {gr+1 , . . . , t} in P3 (where
gr+1 denotes the next downstream node of gr in P3 ) to produce a new path P23 .
Likewise, using the first such node gr in P3 that repeats in P2 (which may be
different from gr ), we can concatenate the nodes {s, . . . , gr } from P3 with
the nodes {gr +1 , . . . , t} in P2 to produce a new path P32 . It is important that
we check the new paths to be sure that they are loop-free. The two offspring
generated in this manner are [P1 , P23 ] and [P32 , P4 ]. On the other hand, if P2
and P3 are disjoint, we could swap P2 with P3 to produce two new offspring
[P1 , P3 ] and [P2 , P4 ].
11.4.5 Mutation
The objective of the mutation operation is to diversify the genes of the current
population, which helps prevent the solution from being trapped in a local
optimum. Just as some malicious mutations could happen in the natural world,
mutation in GA may produce individuals that have worse fitness values. In
such cases, some filtering operation is needed (e.g., the selection operation)
to reject such bad genes and to drive GA toward optimality.
Figure 11.5 An example of the
individual
individual
^
P1 1 3 5 6 9
Original
mutation operation. P1 1 3 5 7 9
New
P2 1 2 5 7 8 9 P2 1 2 5 7 8 9
Mutation is performed on an individual with probability (called the muta-

tion rate). For better performance, we propose a schedule to vary the muta-
tion rate within [min , max ] over iterations (rather than using a fixed ). The
mutation rate is first initialized to max ; then as generation number k increases,
the mutation rate gradually decreases to min , i.e.,
k (max min )
k = max , (11.16)
Tmax
where Tmax is the maximum number of generations. Such schedule of is
similar to the cooling schedule used in SA, and yields better convergence per-
formance for Problem OPT-MM.
Fig. 11.5 illustrates a simple example of the mutation operation. In this
example, we could implement mutation as follows. First, we choose a path
Ph , h = 1 or 2, with equal probabilities. Then, we can randomly pick an inte-
ger value k in the interval [2, |Ph | 1], where |Ph | is the cardinality of Ph ,
and let the partial path {s, . . . , gk } be Phu , where gk is the k-th node along Ph .
Finally, we can use any constructive approach to build a partial path from gk
to t, denoted as Phd , which does not repeat any node in Phu other than gk . If no
such alternative segment exists between gk and t, we keep the path intact; oth-
erwise, a new path can now be created by concatenating the two partial paths
as Phu Phd .
In this section, we present some numerical examples for the GA-based solu-
tion. In each example, we generate a wireless ad hoc network topology by
placing a number of nodes at random locations in a rectangular region, where
connectivity is determined by the distance coverage of each nodes transmitter.
The sourcedestination nodes s and t are uniformly chosen from the nodes.
For each link, the failure probability is uniformly chosen from [0.01, 0.3]; the
available bandwidth is uniformly chosen from [100, 400] Kb/s, with 50 Kb/s
steps; the mean burst length is uniformly chosen from [2,6]. A DD video codec
is implemented and used in the simulations.
We set the GAs parameters as follows: the population size is 15; = 0.7;
is varied from 0.3 to 0.1 using the schedule described in Section 11.4; 2 is
set to 1, since it does not affect path selection decisions. The GA is terminated
after a predefined number of generations or after a prespecified computation
Table 11.2 Comparison of the average distortions obtained

by the GA-based routing and exhaustive search
Topo. 1 Topo. 2 Topo. 3 Topo. 4

Network size 10-node 10-node 15-node 15-node
Global Opt. 0.3308 0.2004 0.3863 0.2969

GA (average) 0.3330 0.2004 0.3937 0.2972
GA (std. dev.) 7.6e-6 0 2.8e-5 2.9e-6
Lower bound 0.2810 0.1832 0.3527 0.2444
time. The best individual found by the GA upon its termination is prescribed
as the solution to Problem OPT-MM.
11.5.1 Near-optimality
One important performance concern is the quality of the GA solutions. As dis-
cussed, due to the complex nature of Problem OPT-MM, a closed-form optimal
solution is not obtainable. However, for small networks, an optimal solution
may be numerically obtained via an exhaustive search and can be used to com-
pare with the proposed GA-based solutions.
Table 11.2 shows the optimal distortion values found by GA (each is the
average of 30 runs) and by exhaustive search for two ten-node and two 15-
node networks. We find that the solutions found by GA are very close to the
global optimum in all cases. In addition, the standard deviation of the 30 GA
results for the same network is very small. The average computational time
for GA is 0.29 s for the ten-node network (about 60 generations) and 0.39
s for the 15-node network (about 70 generations) on a Pentium 4 2.4 GHz
computer (512 MB memory) with MATLAB 6.5. For exhaustive search, the
average computational time is 58.7 s for the ten-node case and 1877 s for the
15-node case. We also compute the lower bound [104] for each of the networks.
The results are given in the last row of Table 11.2.
11.5.2 Comparison with trajectory methods

For comparison purposes, we implemented simulated annealing (SA) and tabu
search (TS) for the same problem. We used the geometric cooling schedule
for the SA implementation with a decay coefficient = 0.99 [1]. For the TS
implementation, we choose a tabu list of five for small networks and ten for
large networks [51].
SA was initially motivated by an analogy between the way a piece of metal
cools and freezes into a minimum energy crystalline structure (annealing pro-
cess), and the search for a minimum in a more general system [1]. When
SA explores the solution space, it accepts a nonimproving solution with a
probability, which decreases with iterations. We use a probabilistic acceptance

function
)
1, if D(x) < D(x)
P r{x x} = (11.17)
exp{ |D(x)D(
ck
x)|
} otherwise,
where ck is a control parameter analogous to temperature in a physical system,

x is the current solution, and x is a perturbation of x. The fashion in which ck
is manipulated is called the cooling schedule. The following cooling schedule
is used in our experiments [1]:
1. c0 = 1: i.e., nearly all transitions will be accepted at the beginning of the

search process;
2. ck+1 = ck : i.e., the control parameter is decremented each time when
a non-improving solution is accepted, and remains at each value for a suf-
ficient time for the system to return to an equilibrium. is the decay
coefficient. We set = 0.99 for all experiments in this chapter.
Compared with SA, TS explicitly uses the history of the search, both to
escape from local minima and to implement an exploratory strategy. Specif-
ically, TS uses a tabu list to prevent returning to recently visited solutions,
therefore avoiding endless cycling and possibly forcing the search process to
accept nonimproving solutions [51]. In our experiments, we use a tabu list of
five for small networks (e.g., ten-node networks) and ten for large networks
(e.g., 50-node networks). The tabu list is implemented using a first-in-first-out
queue. An explored solution is always inserted at the tail of the queue. When
the queue is full, the head of the queue is removed.
In Fig. 11.6, we plot the evolution of distortion values obtained by GA, SA,
and TS for a ten-node network and a 50-node network, respectively. All the
three metaheuristics are terminated after running for 1 s. Upon termination, GA
has evolved 210 generations in Fig. 11.6(a) and 75 generations in Fig. 11.6(b);
SA ran for 1500 iterations in Fig. 11.6(a) and 700 iterations in Fig. 11.6(b); TS
ran for 1050 iterations in Fig. 11.6(a) and 550 generations in Fig. 11.6(b). GA
has fewer number of iterations than SA and TS. For both networks, the best
distortion values found by GA are evidently much better than those by SA or
TS. In Fig. 11.6(a), GA quickly converges to the global optimal, while both SA
and TS are trapped at local optima (i.e., no further decrease in distortion value
after hundreds of iterations). The same trend can be observed in the 50-node
network case shown in Fig. 11.6(b), although the global optimum cannot be
found here. We also plot the lower bounds in the figures.
An interesting observation from Fig. 11.6 is that for GA, the biggest
improvement in distortion is achieved in the initial iterations, while the
improvement gets smaller as GA evolves more generations. The initial pop-
ulation is generated using the random construction method discussed in Sec-
tion 11.4.1, with no consideration of video performance. The initial solutions
Figure 11.6 Comparison of 1

distortion evolution of three 0.9
metaheuristic methods. 0.8
Average distortion
0.7
0.6
0.5 SA GA TS
0.4
0.3
0.2
0.1 Global optimum Lower bound
0
0 0.2 0.4 0.6 0.8 1
Time (second)
(a) Distortion evolution for a 10-node network.
1
0.9
0.8
SA TS
0.7
Average distortion
0.6
0.5
0.4
0.3
0.2 GA Lower bound
0.1
0
0 0.2 0.4 0.6 0.8 1
Time (second)
(b) Distortion evolution for a 50-node network.
usually have high distortion values. The distortion value quickly drops over
iterations, indicating that the GA performance is not very sensitive to the qual-
ity of the initial population. Also note that the SA and TS curves increase at
some time instances (e.g, the TS curve at 0.06 s in Fig. 11.6(a) and the SA
curve at 0.08 s in Fig. 11.6(b)), which implies that a nonimproving solution is
accepted in order to escape from local minima.
In addition to providing better solutions, another strength of GA over trajec-
tory methods is that multiple good solutions can be found after a single run.
Such extra good paths can be used as alternative (or backup) paths if needed.
11.5.3 Comparison with traditional multi-path routings

In this section, we compare the GA-based solution to traditional multi-path
routings, which we call network-centric routing, due to its lack of cross-layer
consideration. We implement two popular network-centric multi-path routing
algorithms, namely k-shortest path (SP) routing (with k = 2 or 2-SP) [38] and
disjoint path routing, Disjoint Pathset Selection Protocol (DPSP) [120]. Our
2-SP implementation uses hop count as the routing metric so that the two
shortest paths are used. In our DPSP implementation, we set the link costs
Table 11.3 Comparison of GA and network-centric routing
Routing P1 Succ. P2 Succ. Desc. 1 Desc. 2 Average

algorithm ratio ratio rate rate PSNR
GA 0.994 0.952 350 Kb/s 350 Kb/s 29.71 dB

2-SP 0.798 0.782 100 Kb/s 200 Kb/s 23.42 dB
DPSP 0.965 0.793 100 Kb/s 100 Kb/s 25.65 dB

to log 1/pij , for all {i, j } E, so that two disjoint paths with the highest
end-to-end success probabilities are used. We compare the performance of our
GA-based multipath routing with these two algorithms over a 50-node ad hoc
network using a video clip.
There are many ways to generate MD video (see [57] for an excellent sur-
vey). We choose a time domain partitioning coding scheme, where two descrip-
tions are generated by separating the even-and odd-numbered frames and cod-
ing them separately. This simple time domain partitioning method is widely
used in many video streaming studies [8; 12; 23; 102]. Compared to a tra-
ditional single description coder, this coder has a comparable computational
complexity. Its coding efficiency is slightly lower than a single description
coder, due to the fact that a longer motion prediction distance is used. However,
this reduced coding efficiency is well justified by the resulting enhanced error
resilience. We use an H.263+ like codec. Since our approach is quite general,
we conjecture that the same trend in performance would be observed for other
video codecs, such as H.264 or MPEG-2 or MPEG-4. The QCIF sequence
Foreman (400 frames) is encoded at 15 fps for each description. A 10%
macroblock level intra-refreshment is used. Each Group of Blocks (GOB) is
carried in a different packet. The received descriptions are decoded and PSNR
values of the reconstructed frames computed. When a GOB is corrupted, the
decoder applies a simple error concealment scheme by copying from the cor-
responding slice in the most recent, correctly received frame.
The quality of the paths found by the three algorithms are presented in
Table 11.3. The 2-SP algorithm has the worst performance in terms of path
success probabilities. The DPSP algorithm has an improved success proba-
bility performance since it uses link success probabilities in routing. How-
ever, it may sacrifice path bandwidth while pursuing low-loss paths. As a
result, it produces the lowest end-to-end bandwidths. We observe that our
GA-based routing yields paths with much higher end-to-end success prob-
abilities and end-to-end bandwidths, resulting in greatly improved video
quality.
The PSNR curves of the received video frames are plotted in Fig. 11.7. We
observe that the PSNR curve obtained by GA is well above those obtained
by the aforementioned network-centric routing approaches. Using GA, the
improvements in average PSNR value over 2-SP and DPSP are 6.29 dB and
Figure 11.7 PSNR curves of 50

received video sequences. 45
40 GA
PSNR (dB)
35
30
25
20
2SP
15
10
5
0
0 50 100 150 200 250 300 350 400
Frame number
(a) GA-based algorithm versus 2-SP.
50
45
40 GA
35
PSNR (dB)
30
25
20
DPSP
15
10
5
0
0 50 100 150 200 250 300 350 400
Frame number
(b) GA-based algorithm versus DPSP.
4.06 dB, respectively. We also experiment with an improved 2-SP algorithm,

which also uses link success probabilities as link metric (as in DPSP). In
this case, our GA-based routing achieves a 1.27 dB improvement over this
enhanced 2-SP version, which is still significant in terms of visual video
quality.
Metaheuristic methods are an important class of heuristic methods and have

been applied to solve some very complex problems in wireless networks. In
this chapter, we gave a review of some well-known metaheuristic methods
(e.g., basic local search, SA, TS, and GA). In the case study, we focused on
developing a GA-based method to solve a multi-path routing problem for MD
video. We found that a GA-based solution is eminently suitable to address this
particular problem, which involves complex objective functions and exponen-
tial solution space. By exploiting the survival-of-the-fittest principle, a GA-
based solution is able to evolve to a population of better solutions after each
iteration and eventually offers a near-optimal solution.
11.7 Problems
11.1 Describe (in greater detail than in Section 11.1) the basic concepts and
approach of the following metaheuristic methods: iterative improvement, sim-
ulated annealing, tabu search, and genetic algorithm. What is the difference
between the genetic algorithm and these other metaheuristic methods?
11.2 Study the so-called Ant Colony Optimization metaheuristics in the
literature, and compare this with the GA approach.
11.3 Why is the set of initial solutions important to the performance of GA?
11.4 Why is the population size important in GA? What is the consequence
if the population size is too large or too small?
11.5 What is the purpose of crossover in a GA? Explain the impact of the
crossover rate on GA.
11.6 What is the purpose of mutation in a GA? Explain the impact of the
mutation rate on GA.
11.7 We use a dynamic k (see Eq. (11.16)) for mutation. What is the advan-
tage of using such a dynamic k ?
11.8 Why is GA eminently suitable for solving a problem such as Problem
OPT-MM?
11.9 In a GA-based algorithm described in this chapter, how can we ensure
that the solutions after crossover and mutation are feasible?
11.10 What are the different randomization steps within GA? Is randomness
important to the genetic algorithm? Why?
PART
IV Other Topics
CHAPTER
12 Asymptotic capacity analysis
Luck is dividend of sweat. The more you sweat, the luckier you get.
Unknown
12.1 Review of asymptotic analysis
In previous chapters, we studied how to optimize a performance metric (e.g.,

a function of throughput) for a given network instance, where the network is
of finite size. In this chapter, we study a different type of problem, where the
goal is to find how much information source nodes can send to their destina-
tion nodes as either the node density goes to infinity or, with the same node
density, the network area goes to infinity. Such a study is called an asymptotic
capacity analysis, also known as capacity scaling law analysis. For such an
investigation, the impact of network topology is usually of little interest and
the focus is on random network topology. This is the approach that we adopt
in this chapter. We consider a random network where each node is randomly
deployed and each node has a randomly chosen destination node. We study an
asymptotic capacity problem under different interference models and derive
related results using the entities (), O(), and (), which are defined as
follows [30]:
f (n) = (g(n)) if f (n) C g(n) for all n > n0 , where C and n0 are pos-
itive constants;
f (n) = O(g(n)) if f (n) C g(n) for all n > n0 , where C and n0 are pos-
itive constants;
f (n) = (g(n)) if C1 g(n) f (n) C2 g(n) for all n > n0 , where C1 ,
C2 , and n0 are positive constants.
283
Table 12.1 A summary of references on asymp-

totic capacity under unicast/multicast/broadcast.
Protocol model Physical model
Unicast [15; 33; 60; 86] [60]

Multicast [92; 138] [63]
Broadcast [61] [62; 91]
Since the seminal work of Gupta and Kumar [60] on capacity scaling laws
for a wireless network, there has been growing interest in this important area.
Related work on random ad hoc networks can be further divided into the
following two categories: unicast capacity (e.g., [15; 33; 60; 86]) and multi-
cast/broadcast capacity (e.g., [61; 62; 63; 91; 92; 138]). In [60], Gupta and
Kumar showed that for a random network, its capacity is ( B ) under
n ln n
the protocol model, and is [( B ), O( Bn )] under the physical model with
n ln n
synchronized power control, where B is the achievable bit rate for a successful
transmission. In [15; 33; 86], the unicast capacity of multi-channel ad hoc net-
works was investigated under the protocol model. Bhandari and Vaidya [15]
studied the capacity of multi-channel single-radio (MC-SR) networks where
there is a set of channels in the network and each node can only switch to a
subset of these channels. Kyasanur and Vaidya [86] studied the unicast capacity
of multi-channel multi-radio (MC-MR) networks where the number of bands
used at a node is limited by the number of radios at the node. Dai et al. [33]
extended the work of [86] to MC-MR networks with consideration of direc-
tional antennas. Keshavarz-Haddad et al. [61] studied the broadcast capacity
for a homogeneous dense network under the protocol model. The broadcast
capacity under the physical model was studied in [62; 91]. Li et al. [92] exam-
ined the multicast capacity of wireless ad hoc networks. Shakkottai et al. [138]
investigated the multicast capacity of large-scale wireless ad hoc networks
under the protocol model. The multicast capacity under the physical model
was analyzed in [63]. We summarize these efforts in Table 12.1.
Finally, there exists some research related to capacity scaling laws under
specialized wireless communication technologies, such as the generalized
physical model [42; 181] (via adaptive modulation and coding scheme), ultra-
wide band (UWB) [177], multiple-input multiple-output (MIMO) [118], net-
work coding [97], multiple-packet reception (MPR) [165], and cognitive radio
networks [75; 162; 176].
12.2 Capacity scaling laws of wireless ad hoc networks
In our case study, we describe how to analyze asymptotic capacities based on

the original work in [42; 60]. We strive for clarity in our presentation in order
285 12.2 Capacity scaling laws of wireless ad hoc networks
to enhance the accessibility of this intricate material to graduate students. To

simplify the discussion, we normalize distance and consider a wireless ad hoc
network comprised of n nodes randomly deployed in a 1 1 area.1 Each node
acts as a source node and transmits data to a randomly chosen destination node
via multi-hop. The per-node throughput (n) is defined as the data rate that can
be transported from each source to its destination. We show how to analyze
the maximum asymptotic per-node throughput (or per-node capacity) as node
density increases.
Since asymptotic capacity analysis depends on the underlying interference
model, we consider the following three interference models in this chapter:
Protocol model: Under this model, a transmission is successful if (i) the

receiver is within a transmission range of the transmitter, and (ii) the receiver
is outside an interference range of other transmitters. The achievable rate of
a successful transmission is a constant B. This model is also called the disk
model and assumes the existence of interference is binary (yes or no).
Physical model: Under this model, a transmission is successful if its signal-
to-interference-and-noise ratio (SINR) is over certain threshold . The
achievable rate of a successful transmission is assumed to be a constant B.
This model corresponds to a coding scheme that can support transmission
with a data rate B if the SINR threshold is satisfied.
Generalized physical model: Under this model, the achievable rate B of
a transmission is determined by the Shannon capacity formula, i.e., B =
W log2 (1 + SINR). Unlike the physical model, there is no SINR threshold
for a successful transmission. As long as the transmission rate is no more
than W log2 (1 + SINR), such a transmission rate can be supported.
We can see that the protocol model is the simplest model, while the gener-
alized physical model is the most challenging one. In the rest of this chapter,
we study capacity scaling laws under these three interference models. For each
model, we will develop a feasible solution, which will be used as a capacity
lower bound. We will also analyze a capacity upper bound. The results for
these three models are summarized as follows:
Case 1: Under the protocol model, the capacity of a wireless ad hoc network
with n nodes is ( B ) almost surely as n . By almost surely,
n ln n
we mean that the probability of the event that the capacity is ( B )
n ln n
approaches one as n .
1 In [60], to avoid the boundary effect, it is assumed that nodes are deployed on the surface of a
three-dimensional sphere. As we shall see in this chapter, such an assumption is not needed.
Case 2: Under the physical model, the capacity of a wireless ad hoc net-
work with n nodes is (n) [( B ), O( n1/B
)] almost surely as n ,
n ln n
where is the path loss index.
Case 3: Under the generalized physical model, the capacity of a wireless ad
hoc network with n nodes is ( Bn ) almost surely as n .
In Sections 12.3, 12.4, and 12.5, we give details on how to develop these
results. Section 12.6 summarizes this chapter. Table 12.2 lists the notation used
in this chapter.
Table 12.2 Notation.
General notation
a(n) The size of a small square cell

B The maximum data rate from a transmitter to a receiver
dij The distance between nodes i and j
E(t) The set of active links at time t
l(n) The side-length of a cell
n The number of nodes in the network
p(n) The common transmission power at all nodes under synchronized power
control or the protocol model
P r{A} The probability of an event A
SINRij The SINR on link (i, j )
Path-loss index
Ambient noise power
(n) Per-node throughput of a wireless ad hoc network
A feasible solution
Protocol model-specific notation
D The mean distance between a source node and its destination node
r(n) The common transmission range for all nodes
A parameter to set the interference range
Physical model-specific notation
d(q, b) The length of the q-th hop traversed by the bth unit of data
gij The channel gain on link (i, j )
h(b) The number of hops traversed by the b-th unit of data from its source
to the destination

H = nT b=1 h(b), i.e, the total number of hops traversed by all data in
a time duration T
pij (t) The transmission power used by node i to transmit to node j at time t
under an independent power control
An SINR threshold for successful transmissions
Generalized physical model-specific notation
Bm A grid of m(n) horizontal

and m(n) vertical paths

Bwh A box with width w l(n) 2 and height h l(n) 2
287 12.3 Case 1: Asymptotic capacity under the protocol model
Table 12.2 (cont.)

Bwh The event that there is a crossing path between Bwh s left and right
sides
cij (d) The data rate that a node i in a cell can transmit to a node j in another
cell at most d (diagonal) cells away
m(n) The number of cells crossed by a horizontal or vertical path
P The common transmission power at all nodes in the feasible solution
Pmi The maximal number of edge-disjoint crossing paths of rectangle Rm i
between its left and right sides
q The probability that there is at least one node in a cell
Rmi The i-th rectangle of size 1 ln m(n)
m(n)
Swh The dual box of Bwh

Swh The event that there is a crossing path between Swh s top and bottom
sides
W Channel bandwidth
12.3 Case 1: Asymptotic capacity under the protocol model
In this section, we analyze the capacity scaling law under the protocol model.
Under this model, a node i can successfully transmit data with a rate B to a
node j if and only if the following constraints hold:
The receiving node j is within the transmission range of node i, i.e.,
dij r(n),
where dij is the distance between nodes i and j and r(n) is the transmis-
sion range. The above constraint implicitly sets a constraint on next-hop
routing, i.e., it defines a set of possible candidate nodes (within the trans-
mission range) as the next-hop node.
For any other link (k, l) that is active at the same time, it is necessary that
dkj (1 + )r(n), where (1 + )r(n) represents an interference range.
This is to keep the concurrently transmitting node k sufficiently far away
from node j so that ks interference on j is negligible. More formally, we
have
dkj (1 + )r(n) for each link (k, l) E(t) and (k, l) = (i, j ), (12.1)
where E(t) is the set of links in the network that are active at time t. Con-
straint (12.1) implicitly sets a constraint on scheduling, i.e., it defines conflict
relationships among all the links in the network.
Note that under the protocol model, we assume that the same (synchronized)
transmission power p(n) is used for all nodes. Thus, the transmission range
r(n) is the same for all nodes.
We consider a common throughput (n) for each node (source) in the net-
work to its randomly selected destination. The goal is to find the maximum
(n) that can be transported by the network. In Section 12.3.1, we develop a
capacity upper bound. The lower bound analysis is given in Section 12.3.2.
The results of this section can be summarized as follows (Theorems 12.1 and
12.2):
Under the protocol model, the capacity of a wireless ad hoc network with n
nodes is (n) = ( B ) almost surely as n .
n ln n
12.3.1 A capacity upper bound

For a wireless ad hoc network, the required resource to transmit data from a
source node to its destination node depends on the number of hops between
these two nodes, which is yet to be determined. As a result, it is not easy
to analyze the capacity upper bound directly. Instead of analyzing capacity
directly, we can analyze the sum of transmission rates over all nodes in the
network, which we call an aggregate rate (AR). We find that AR should be no
less than a minimum required value, which is a function of (n). We also find
that AR is upper bounded by a constant. Combining these two results, we can
compute an upper bound on throughput (n). The details are given in the proof
of Theorem 12.1.
Theorem 12.1
Under the protocol model, a capacity upper bound is O( B ) almost
n ln n
surely as n .
Proof. We begin the proof by analyzing the minimum required AR in order to

support a throughput (n) at each node. Let D be the mean distance between a
source node and its corresponding destination node. Note that under the proto-
col model, the mean number of hops for each sourcedestination pair is at least
D D
r(n) and there are n sourcedestination pairs. Thus, AR is at least r(n) n(n).
We now develop an upper bound for AR. Based on the scheduling constraint
(12.1), for any two links (i, j ) and (k, l) that are active on the same band
at the same time, we have dik dj k dij (1 + )r(n) r(n) = r(n),
i.e., the distance between any two concurrent transmitting nodes is at least
r(n). Thus, if we draw a disk of radius r(n)
2 at each of these transmitting
2 2
nodes, then these disks must be disjoint. Since the area of each disk is 4r (n)
and the center of these disks are within a unit square, the number of these disks
2
is at most (1+r(n))
2 r 2 (n)/4
= O(r 2 (n)). Thus, the network can support O(r 2 (n))
transmissions at any time. Therefore, summing the data rates of all the trans-
missions, and considering the fact that the rate of a transmitting node is B, we
have that AR is upper bounded by O( r 2B(n) ).
Combining the above two results, we have D

r(n) n(n) AR O( r 2B(n) ).
Note that D is a constant. Therefore, we have (n) = O( nr(n)
B
). It has been
0
shown in [59] that the transmission range r(n) must be greater than ln nn to
make the network connected almost surely as n . Therefore, we have

% &
B B B
(n) = O( )<O 0 =O
nr(n) n ln n n ln n
n
almost surely as n .
12.3.2 A constructive lower bound

To develop a capacity lower bound, it is sufficient to construct a feasible solu-
tion that satisfies all the constraints. A solution contains two parts: a routing
scheme and a scheduling scheme. We consider a cell-based approach for the
routing scheme. In particular, we divide the unit square area into small cells
so that there is at least one node in each cell almost surely as n (see
Lemma 12.1). For each sourcedestination pair, the straight line that connects
them will pass through a number of cells. We choose a node from each of
these cells as a relay for multi-hop routing. To avoid interference, scheduling
must be performed along with cell-based routing. That is, under the protocol
model, some links cannot be active at the same time. Thus, we need to arrange
them into different time slots. To analyze the achieved throughput, we need
to identify the number of time slots required for scheduling. We will show
that the scheduling problem can be transformed into a vertex-color problem in
graph theory. By applying a result from graph coloring, we know how many
time slots will be used for scheduling. Consequently, we obtain a complete
solution, the throughput of which is stated in the following theorem:
Theorem 12.2
Under the protocol model, we can construct a feasible solution with a
throughput of (n) = ( B ) almost surely as n .
n ln n
We now give details for the routing and scheduling schemes described in the
above discussion.
Routing scheme We develop a cell-based routing

0 scheme. We divide the
unit square into small square cells with width lnnn (see Fig. 12.1). The
Figure 12.1 Multi-hop routing a(n)

from a source node s to its
destination node d.
S
0 0
area of a cell is a(n) = lnnn lnnn = lnnn .2 We set the transmission range

r(n) = 5a(n) so that a node in one cell can transmit to a node in any of its
four neighboring cells. Next, we draw a line to connect each sourcedestination
pair, which passes through some cells. One node is chosen from each of these
cells to relay the traffic from the source node to its destination (see Fig. 12.1
as an example).
Such a routing scheme requires at least one node in each cell. We call a cell
without any node an empty cell. We have the following result regarding the
nonexistence of an empty cell as n :
Lemma 12.1
If the cell size is a(n) = lnnn , an empty cell is nonexistent almost surely as
n . Therefore, the routing scheme is feasible almost surely as n .
a(n) = ln n cells, each node can be in any cell with

1 n
Proof. Since there are
an equal probability of lnnn . For a particular cell i, denote Ei as the event
that this cell is an empty cell. Then the probability of Ei is P r(Ei ) = (1

n ) . Therefore, the probability that an empty cell exists is P r( i Ei )
ln n n

i P r(Ei ) = ln n (1 n ) . As n , this probability is
n ln n n
2 To divide the unit square into an integral number of small cells, we need to set
a(n) = 1/n/ ln n. However, to make the proofs in this section easy to understand, we
neglect such integrality requirements and simply let a(n) = 1/(n/ ln n) = ln n/n. Note that all
the results in this section hold true if we set a(n) = 1/n/ ln n. For example, we ask the
readers to prove in Exercise 6.10 that Lemma 12.1 still holds when a(n) = 1/n/ ln n.
Figure 12.2 The shaded area

for Xi . A
D B
Xi
% & % & n
n ln n n n ln n ln n ( ln n)
lim 1 = lim 1
n ln n n n ln n n
n ln n
= lim e
n ln n
1
= lim
n ln n
= 0,
where the second equality holds because limn n

ln n and limx
(1 x1 )x = e.
For this routing scheme, we obtain the following result for the number of
S-D lines passing through any cell:
Lemma 12.2

The number of S-D lines passing through any cell is O(n a(n)) almost
surely as n .
To prove this lemma, we need an intermediate result as follows. Denote Li ,

1 i n, as the n S-D lines. Denote ci , 1 i m = lnnn , as the m small
cells. Define pij as the probability that the S-D line Li passes through cell Qi .
Then, we have the following result regarding pij :
Lemma 12.3
For any 10 i n and 1 j n, there exists a constant c1 > 0 such that
pij c1 ln n
n .
0
Proof. Note that cell Qj is contained in a disk of radius dr = 1 l(n) = ln2nn
2
that is centered at Qj s center D (see Fig. 12.2). Denote Xi and Yi as the
source and destination nodes of the S-D pair i, respectively. Note that Xi can
fall either inside or outside the disk.
We first consider the scenario when Xi falls outside the disk. Suppose that
Xi is at distance x from the disk. We extend the two tangent
lines origination
from Xi equally such that |Xi A| = |Xi B| and Xi C = 2, where C is the mid-
point of line AB. Then Li passes through Qj only if Yi is in the shaded area.
Its area is less than the minimum of 1 (the area of the unit square) and the area
of the triangle, which is equal to 2 2 2 < 2dx r .
(x+dr ) dr
Since Xi is uniformly distributed, the probability density that it is at a dis-
tance x away from the disk is upper bounded by c2 (x + dr ) for some constant
c2 > 0. Thus, we have
P r {Li passes through cell Qi |Xi is outside the disk}

7 2 ' <
2dr
min 1, c2 (x + dr )dx
0 x
7 2dr 7 2
2dr
= c2 (x + dr )dx + c2 (x + dr )dx
0 2dr x
=
ln n
c3 , (12.2)
n
where c3 is a constant.
For the case when Xi falls inside the disk, we have
P r {Li passes through cell Qi |Xi is inside the disk} 1. (12.3)
Combining (12.2) and (12.3), we have
pij = P r {Li passes through cell Qi |Xi is outside the disk}

P r {Xi is outside the disk}
+ P r {Li passes through cell Qi |Xi is inside the disk}
P r {Xi is inside the disk}
=
ln n
c3 (1 rd2 ) + 1 rd2
n
=
ln n
c4 .
n
We are now ready to prove Lemma 12.2.
Proof of Lemma 12.2. Denote

'
1 if Li passes through cell Qj ,
Yij =
0 otherwise,
for 1 i n and 1 j m. Then, we have P r{Yij = 1} = pij . Note that

for a fixed j , the Yij -variables are i.i.d.

Denote Zj = ni=1 Yij as the total number of S-D lines passing through cell
Qj . Consider a fixed j , i.e., a particular cell Qj . For any s, a > 0, based on
the Chernoff bound, we have
E[eaYj ]
P r{Zj > s} . (12.4)
eas
For E[eaZj ], we have
n >
n >
n
E[eaZj ] = E[ea i=1 Yij ]= E[eaYij ] = (ea pij + e0 (1 pij ))
i=1 i=1
>
n
= (1 + (ea 1)pij )
i=1
- .n
=
ln n
1 + (e 1)c4
a
n
- # = $.n
ln n
exp (ea 1)c4
n

= exp(c4 (ea 1) n ln n),
the last inequality holds due to the fact that 1 + x e . Letting s =

where x
c5 n ln n in (12.4) (where c5 > c4 ), we have

P r{Zj > c5 n ln n} exp[ n ln n(c4 (ea 1) ac5 )].
The above inequality holds for any 1 j m.

Since c5 > c4 , we can let a be small enough so that

P r{Zj > c5 n ln n} exp( n ln n),
for some constant > 0. Then, by union bounds, we have

P r{there exists a cell having more than c5 n ln n S-D lines passing through}

m
P r{Zj > c5 n ln n}
j =1
n
m exp( n ln n) = exp( n ln n),
ln n
which goes to zero as n .
Scheduling scheme We consider a time-slot-based scheduling, i.e., we

divide one time frame into multiple time slots to satisfy the protocol model
scheduling constraint (12.1).
To analyze the performance of the scheduling scheme, we need to know
the number of conflicting links for each link (i, j ), which directly affects the
number of time slots required for scheduling and throughput. Two links are in
conflict if they cannot be active at the same time under the protocol model. We
Figure 12.3 The area that

covers all the transmitting nodes
that may interfere with the
receivers in cell Q.
(1 + )r(n) a(n) (1 + )r(n)
now establish the following lemma on the number of conflicting links for any
link in the network:
Lemma 12.4
Under the cell-based routing scheme, the number of conflicting links for

any link is upper bounded by O(n a(n)) almost surely as n .
Proof. Note that for a link (i, j ), there are two cases in which a link (k, l) could
be its conflicting link. One case is that link (k, l) interferes with link (i, j ) and
the other case is that link (k, l) is interfered by link (i, j ). In this proof, for a
given link (i, j ), we first analyze the number of conflicting links that interfere
with link (i, j ). Since the routing scheme is cell-based, the conflict analysis is
also based on cells. That is, we first analyze the number of interfering cells and
then analyze the number of conflicting links in an interfering cell. We say that
a cell Q is an interfering cell with respect to cell Q if and only if the distance
between a transmitting node in cell Q and a receiving node in cell Q is no
more than (1 + )r(n).
We now show that the number of interfering cells with respect to cell Q
is a constant. For a receiving node j in cell Q, the transmitting nodes of the
links that interfere with j must be within the area inside the solid line shown in

Fig. 12.3, which consists of five rectangles (including the small cell a(n)

a(n) at the center) and four quarter disks with radius (1 + )r(n). To obtain
an upper bound, we define the outermost square area as the interfering area that
will not have any transmitting node in it. Hence, the number of interfering cells
is no more than c6 = (2 (1+)r(n)

a(n)
+ 1)2 = (2 5(1 + ) + 1)2 , which is
a constant.
Next, we analyze the number of links that interfere with link (i, j ) in
each interfering cell. Clearly, this number is no more than the number of
transmissions in this cell. Based on Lemma 12.2 and the routing scheme
discussed earlier, the number of transmissions in a cell is equal to the number
295 12.4 Case 2: Asymptotic capacity under the physical model

of sourcedestination lines intersecting this cell, which is O(n a(n)) almost
surely as n .
Combining the above results, the number of conflicting links that interfere

with link (i, j ) is given by c6 O(n a(n)) = O(n a(n)) almost surely as
n .
Following a similar analysis, we can obtain the same result on the
number of conflicting links that are interfered by link (i, j ). Then the
number of all conflicting links for a link (i, j ) is upper bounded by

O(n a(n)) + O(n a(n)) = O(n a(n)) almost surely as n .
To schedule the conflicting links, we use a conflict graph to model them.

Each link in the network corresponds to a vertex in the conflict graph, and any
conflict in the network is represented by an edge connecting two corresponding
vertices in the conflict graph. If we use a different vertex color to represent
each time slot, then the scheduling problem reduces to the well-studied vertex-
color problem, which aims to find the minimum number of colors needed for
coloring the nodes in a graph such that any two nodes connected by an edge
have different colors. Using Lemma 12.4, the degree of each vertex in the

conflict graph will be at most c7 n a(n) for some constant c7 almost surely
as n . Note that by the result for graph coloring [168], it is enough to
have d + 1 colors for a graph with the maximum degree d. Thus, the required

number of colors for the conflict graph is at most 1 + c7 n a(n). So we can

divide one time frame into at most 1 + c7 n a(n) equal length time slots for
scheduling. Therefore, the achievable throughput (n) is given by
% &
B B B
(n) = = ,
1 + c7 n a(n) 1 + c7 n ln n n ln n
where the first equality holds because a(n) = ln n
n .
12.4 Case 2: Asymptotic capacity under the physical model
In this section, we analyze the capacity scaling law under the physical model.
Under this model, each node is allowed to perform power control. Denote
pij (t) as the power used by node i to transmit to node j .3 A transmission
with a data rate B is successful if and only if the SINR satisfies
gij pij (t)
SINRij = (k,l)=(i,j ) , (12.5)
+ (k,l)E(t) gkj pkl (t)
3 In [60], the transmission power at all nodes in a random network is the same and is determined
by the number of nodes. A different capacity upper bound O( 1/2 B ) was obtained for this
n
special case. In this section, we study a general case where independent power control is
allowed at each node.
where gij = dij is the channel gain over link (i, j ), 2 is the path-loss
index, is the ambient noise power, and is the SINR threshold.
In Sections 12.4.1 and 12.4.2, we develop an upper bound and a lower
bound, respectively. The main result can be summarized as follows (also see
Theorems 12.3 and 12.4):
Under the physical model, the capacity of a wireless ad hoc network with n
nodes is (n) [( B ), O( n1/
B
)] almost surely as n .
n ln n
12.4.1 Computing an upper bound

Instead of analyzing capacity directly, we analyze the total distance traversed
by all data generated over a time duration T (i.e., (n)nT ), which we call
aggregate capacity-distance over T , denoted as ACD T . Note that ACD T is
similar to the transport capacity in [60]. We find that there exists an inequality
relationship between ACD T and the throughput (n) and that a constant upper
bound can be computed for ACD T . As a result, we can compute an upper
bound for the throughput (n).
Theorem 12.3
Under the physical model, the capacity of a wireless ad hoc network with n
nodes is (n) = O( n1/
B
) almost surely as n .
Proof. We begin the proof by finding an inequality relationship between

ACD T and the throughput (n). During T , the network can transport (n)nT
units of data in total. For a particular unit of data b, denote h(b) as the number
of hops on the path from its source to its destination and let d(q, b) denote the
length of the q-th hop. Then we have

(n)nT h(b)
ACD T = d(q, b) (n)nT D, (12.6)
b=1 q=1
where D is the average distance between the source and destination nodes.
Next, we develop an upper bound for ACD T . Since the function f (x) = x
is convex for any 2, we have

(n)nT h(b)
1 1
(n)nT h(b)
d(q, b) d (q, b),
H H
b=1 q=1 b=1 q=1

where H = nT b=1 h(b) is the total number of hops traversed by all the data
over time duration T . This gives us

(n)nT h(b)
ACD T = d(q, b)
b=1 q=1
1

1
(n)nT h(b)
H d (q, b)

H
b=1 q=1
1

(n)nT h(b)
=H 1 1 d (q, b) .

(12.7)
b=1 q=1
h(b)
We now analyze H and (n)nT b=1 q=1 d (q, b). We first obtain an upper
bound for H . Note that due to half-duplex, a node that is receiving cannot
transmit at the same time. Therefore, in any given time slot, at most n2 nodes
can transmit simultaneously. For each link (i, j ), its capacity is B. Thus, the
total bits that can be transmitted by all nodes over T is at most nBT
2 , i.e.,
nBT
H . (12.8)
2
(n)nT h(b)
We next find an upper bound for b=1 q=1 d (q, b). Consider a trans-
mission from node i to node j . Since the SINR constraint (12.5) holds at node
j , we have
pij (t)/dij
(k,l)=(i,j ) .
+ (k,l)E(t) pkl (t)/dkj

Adding the signal power from node i in the denominator, we have
pij (t)/dij
.
+
(k,l)E(t) pkl (t)/dkj +1
Based on this observation, we can develop an upper bound on dij as follows:
+1 pij (t)
dij pkl (t)
+ (k,l)E(t)
dkj
+1 pij (t)
p
kl (t)
+ (k,l)E(t)
( 2)
+1 pij (t)
< p
,
kl (t)
(k,l)E(t) ( 2)

where the second inequality holds because dkj 2. Summing the above
inequality over all simultaneous transmissions, we have

+1 (i,j )E(t) pij (t) +1

dij < p (t)
= ( 2) .
(k,l)E(t)
kl
(i,j )E(t) ( 2)
Further, summing the above inequality over a time duration T , we have

T
T
+1 +1
dij < ( 2) = T ( 2) . (12.9)

t=1 (i,j )E(t) t=1
In addition, we have

(n)nT h(b)
T
T
d (q, b) = (dij B) = B dij , (12.10)
b=1 q=1 t=1 (i,j )E(t) t=1 (i,j )E(t)
where the first equality holds since each active link (i, j ) can carry B bits of
data in a unit time. Combining (12.9) and (12.10), we have

(n)nT h(b)
+1
d (q, b) < BT ( 2) . (12.11)

b=1 q=1
By (12.7), (12.8), and (12.11), we have

% & 1 1 1
nBT 1 +1 nBT 2( + 1)
ACD T < BT ( 2) = .
2 2 n
(12.12)
Now we are ready to develop an upper bound for (n). Based on (12.6) and
(12.12), we have
1
nBT 2( + 1)
(n)nT D ACD T < .
2 n
Thus,
1 % &
B 2( + 1) B
(n) < =O
2D n n1/
almost surely as n .
12.4.2 Computing a lower bound

To obtain a capacity lower bound, we construct a feasible solution. But con-
structing a feasible solution under the physical model is more difficult than that
under the protocol model. This is because under the protocol model, we only
need to consider distances in the scheduling constraints (12.1). But under the
physical model, we also need to consider power control in the SINR constraints
(12.5).
Given that we have developed a feasible solution for the protocol model,
we will try to develop a feasible solution for the physical model. We observe
that if we set the parameter in the protocol model to be large enough, then
is also a feasible solution for the physical model. This is because a large
will impose more constraints within the conflict graph, and thus reduce inter-
ference from neighboring nodes in the protocol model solution. As a result,
the SINR at a receiver can be made large enough to satisfy the physical model
SINR constraint (12.5). In this case, is also a feasible solution for the phys-
ical model. The following theorem is based on this insight:
Theorem 12.4
Under the physical model, we can construct a feasible solution with
(n) = ( B ) almost surely as n .
n ln n
Proof. To prove this result, we show that if we set

% & 1
p(n) 2+
1, (12.13)
2r 2 (n)(p(n) r(n) )
where p(n) is the transmission power used in the protocol model solution ,
then the solution developed for the protocol model (in Section 12.3.2) is also
feasible for the physical model.
First, we show that the SINR constraint under the physical model (12.5) is
satisfied by . Based on the construction of in Section 12.3.2, we know
that once a link (i, j ) is active, those nodes that are within a square with a

side length of 2(1 + )r(n) + a(n) (that contains the cell where node j is
located) cannot transmit at the same time (see Fig. 12.3). Note that node j may
be located at the boundary of the unit square. Thus, the number of links that
interfere with link (i, j ) is at most
? @2 % &2
2(1 + )r(n) + 1 2(1 + )r(n) + 1
1 1
2(1 + )r(n) + a(n) 2(1 + )r(n) + a(n)

[1 a(n)][4(1+)r(n)+ a(n)+1]
=
[2(1 + )r(n) + a(n)]2
2
<
[2(1 + )r(n) + a(n)]2
2
<
4[(1 + )r(n)]2
1
= ,
2[(1 + )r(n)]2

where the second inequality holds because 1 a(n) < 1 and 4(1 +

)r(n) + a(n) 0 < 1 as n . Based on the construction of , the
interference at the receiving node j from each of these links will be at most
p(n)
[(1+)r(n)] . Therefore, we have
p(n)
r(n)
SINRij > p(n)
+ 1
2[(1+)r(n)]2 [(1+)r(n)]
r(n)
=
p(n)1 + 12 [(1 + )r(n)](2+)
r(n)
= ,
p(n)1 + [ 1 r(n) p(n)1 ]
where the second inequality holds by (12.13). Thus, for solution , the SINR
constraint (12.5) holds for each link (i, j ) that is active at time t. There-
fore, the power control and scheduling schemes in are also feasible for
the physical model. Furthermore, the achieved capacity on each link under
the physical model is equal to that under the protocol model. As a result,
the link capacity constraint is still satisfied for each link under the physi-
cal model, and the achieved throughput under the physical model will be the
same as that under the protocol model, which is ( B ) almost surely as
n ln n
n .
12.5 Case 3: Asymptotic capacity lower bound under the

generalized physical model
In this section, we analyze the capacity scaling law under the generalized phys-
ical model. Under this model, the achievable rate of a transmission is deter-
mined by its SINR. That is, the achievable rate from node i to node j is
cij = W log2 (1 + SINRij ).
Under this model, the only requirement is that the transported data rate is no
more than this achievable rate cij .
This model is the most challenging one among the three models. It remains
an open problem to compute an asymptotic upper bound for this model. In this
section, we present a capacity lower bound by applying the percolation theory
based on the materials in [58; 107]. The following theorem is the main result
in this section:
Theorem 12.5
Under the generalized physical model, the capacity of a wireless ad hoc
network generated by a Poisson point process with density n in a 1 1 area
is ( Wn ) almost surely as n , where W is the channel bandwidth.
301 12.5 Case 3: Asymptotic capacity lower bound under the generalized physical model
Remark 12.1
The above result was established in [42]. Based on this result, the authors
of [42] claimed that this result closes the capacity gap for random networks
since it matches with the upper bound developed in [60]. Unfortunately, this
claim is incorrect for the following reasons:
The underlying interference models are different. In [60], the upper

bound O( Wn ) was developed for the physical model, while in [42], the
lower bound ( Wn ) was developed for the generalized physical model.
We cannot assume (and it is usually incorrect to claim) that the same
capacity upper bounds can hold under different interference models.
The random networks under consideration are also different. In [42], it
was assumed that each node is the destination of exactly one source, while
there is no such assumption in [60]. We also point out that the number of
nodes in the random network in [42] may not be n, while the number of
nodes in the random network in [60] is n. Given this difference in network
settings, we cannot assume that the upper bound result in [60] also holds
for the random network in [42].
12.5.1 Main idea

We consider a random network generated by a Poisson point process with den-
sity n in a 1 1 area.4 We assume that each node is the destination of exactly
one source. To obtain a capacity lower bound, we assume that all nodes trans-
mit at the same constant power P . We also divide the entire network area into
small cells. Our solution on multihop routing is based on a highway system
in the network. A single node in each cell that is crossed by a highway path
is selected to transmit data along this highway. The other nodes transmit to,
or receive from, the highway system via a single hop. That is, an end-to-end
transmission has four phases: (i) nodes send their data to a node in the high-
way via one-hop transmissions; (ii) data are carried by a horizontal highway
path; (iii) data are carried by a vertical highway path; and (iv) data are delivered
from a node in the highway to the destination nodes via one-hop transmissions.
Note that if the source node (or destination node) is in the highway, then Phase
(i) (or Phase (iv)) is not needed. In each phase, we design a time-slot-based
scheduling for transmissions. Since Phases (i) and (iv) have longer transmis-
sion distances (and thus a lower peak data rate), there are more time slots to
ensure an appropriate data rate in these phases.
4 In previous sections, the number of nodes in the network is n. In this section, the number of
nodes in the network is a random number with an expectation value of n.
Figure 12.4 Construction of the l(n) 2

grid.
l(n)
si
(a) Cells. (b) The grid.
We now provide a sketch of the results. First, we consider the data rate limit
in Phases (ii) and (iii). The constructed highway system has hops with a con-
stant transmission rate. We will also show that the number of nodes that have

access to a highway path is at most O( n). Thus, each source node has a data

rate no more than (W/ n), where W is the channel bandwidth. Second, we
consider the data rate limit in Phases (i) and (iv). We show that the length of

these single hops is at most O(1/ n), which in turn yields a rate ( (lnWn)3 ).
Combining both results, we establish Theorem 12.5.
12.5.2 Construction of the highway

0
To construct the highway, we select an integer m(n) = 2 ln
n
6 and then par-
tition the area into small cells si of side-length l(n) = 1
, as depicted in
m(n) 2
Fig. 12.4(a). Denote X(si ) as the number of nodes inside cell si . If X(si ) = 0,
then we say that cell si is empty. Each cell is nonempty with the same
probability q P r{X(si ) 1}, where P r{A} is the probability of an event A.
n
n 2 ln 6
We have q = 1 e(l(n)) n = 1 e 2(m(n))2 > 1 e 2 n = 56 . Moreover,
2
since q 56 as n , we have q < 1 6e13 as n . Thus,
5 1
< q < 1 3 when n . (12.14)
6 6e
For the number of nodes in a cell, we have the following lemma:
Lemma 12.5
There are less than ln m(n) nodes in each cell almost surely as n .
To prove this lemma, we need the following result:
Lemma 12.6
For a Poisson random variable X with parameter , we have
e (e)x
P r(X x) for x > .
xx
Proof. For s > 0 and x > , based on Markovs inequality, we have

E(esX )
P r(X x) , (12.15)
esx
where E(esX ) can be computed as follows:
k k
e
E(esX ) = esk
k!
k=0
es

(es 1) e (es )k
=e
k!
k=0
s 1)
= e(e . (12.16)
Combining (12.15) and (12.16), we obtain
E(esX )
= e(e 1)sx .
s
P r(X x)
esx
Letting s = ln x > 0 in the above inequality, we have
e (e)x
P r(X x) exx ln(x/) =
xx
for x > .
We are now ready to prove Lemma 12.5.
Proof of Lemma 12.5. We compute the probability P r that there is at least

one cell that has no fewer than ln m(n) nodes. By the union bounds and
Lemma 12.6, we have
P r [m(n)]2 P r{X(si ) ln m(n)}
ln m(n)
(l(n))2 ne
[m(n)]2 e(l(n)) n
2
ln m(n)
ln m(n)
n ne3
= e 2(m(n))2
2(m(n))2 ln m(n)
= e(1) (ln n) ln n ,
which goes to 0 as n .
Figure 12.5 There are many 1

paths between left and right
i . k ln m(n) R1m
sides in each rectangle Rm
m(n)
R2m
m(n)
Rmk ln m(n)
We draw m(n) horizontal lines and m(n) vertical lines across half of the
cells, as shown in Fig. 12.4(b). A path includes some segments of these lines. If
a path does not cross any empty cell, then we call it an open path. We will show

that there are at least ( n) open paths crossing the network area between left

and right sides and at least ( n) crossing paths between top and bottom sides
almost surely as n . We call these paths the highway system. Along the
paths of the highway system, we select one node per cell that acts as a relay.
We now present a proof for crossing paths between left and right sides. The
proof for crossing paths between top and bottom sides is similar and is thus
omitted. We divide the network area into rectangles Rm i of size 1 ln m(n)
m(n)
(see Fig. 12.5), where the constant is given by = * m(n) +. We
m(n) ln(66q)
ln m(n) 3 ln m(n)
note that
m(n) 3
> = > 1, (12.17)
ln m(n) m(n) ln(66q)
3 ln m(n)
ln(6 6q)
where the last equality holds because q < 1 6e13 by (12.14). Denote Pmi as
the maximal number of edge-disjoint crossing paths
(between left andright
m(n)
sides) within the rectangle Rm , and let Pm = min Pm : 1 i ln
i i
m(n) . We
have the following result:
Theorem 12.6
ln(66q)3
For a constant = q
ln 1q
> 0, we have that
lim P rq {Pm ln m(n)} = 0,

n
where P rq {A}
0 is the probability of event A under a nonempty probability
q, m(n) = * m(n)
2 ln 6 , and =
n + . That is, there are at least
m(n) ln(66q)
ln m(n) 3 ln m(n)
ln m(n) + 1 edge-disjoint crossing paths within each rectangle Rm

i of
ln m(n)
size 1 m(n) between left and right sides almost surely as n .
In Theorem 12.6, we can now better understand why we set =

* m(n) + . Since should be a positive number, we need >
m(n) ln(66q)
ln m(n) 3 ln m(n)
3 m(n)
ln(66q) .
Furthermore, the number of rectangles, ln m(n) , should be an inte-
ger. The that we choose satisfies both requirements.
m(n)
Note that based on Theorem 12.6, it is clear that there are at least ln
m(n)
( ln m(n) + 1) = (m(n)) = ( n) crossing paths in the network area
between left and right sides almost surely as n .
The proof of Theorem 12.6 needs some additional lemmas. We denote by

Bwh the event that there is a crossing path within Bwh between its left and

right sides, where Bwh is a box with width w l(n) 2 and height h l(n) 2.
The following lemma provides an upper bound for the probability of the event
:
Bwh
Lemma 12.7
} of the event
For a nonempty probability q < 13 , the probability P rq {Bwh

Bwh satisfies

P rq {Bwh } (h + 1)eln(3q)w .
Proof. Denote Ai as the event that there exists a crossing path starting from
the i-th vertex on the left side of box Bwh , where vertices on the left side are
}.
ordered from the bottom to the top. Thus, we have i P rq {Ai } = P rq {Bwh
Since the number of vertices is at most h + 1, there is at least one index i0
such that P rq {Ai0 } h+1
1 }, i.e.,
P rq {Bwh

P rq {Bwh } (h + 1)P rq {Ai0 }.
Denote Ni0 (w) as the number of open paths with w hops starting at the
i0 -th vertex. Since a crossing path starting from the i0 -th vertex to the right
side of Bwh has at least w hops, we have
P rq {Ai0 } P rq {Ni0 (w) 1}.
Note that paths with w hops are open with probability q w and the number
of paths with w hops starting at the i0 -th vertex is at most 3w . We have
P rq {Ni0 (w) 1} q w 3w = (3q)w

Figure 12.6 The grid Bm (with

continuous lines) and its dual
grid Sm (with dashed lines).
almost surely for a large number of open paths. Combining the above inequal-
ities, we have

P rq {Bwh } (h + 1)P rq {Ai0 }
(h + 1)P rq {Ni0 (w) 1}
(h + 1)(3q)w
= (h + 1)eln(3q)w
for a nonempty probability q < 13 .
The above lemma asserts that if the nonempty probability is too small, then
a box Bwh cannot be crossed between its left and right sides. We consider
the dual question of the existence of a crossing path by exploiting the concept
of a dual grid from the Percolation theory. A dual grid Sm for the grid Bm is
constructed by placing a vertex in each cell of Bm , and connecting two neigh-
boring vertices by an dashed line, as shown in Fig. 12.6. For each edge in Sm
that does not cross an open edge of the original gird, we call it an open edge
in Sm . We have the following relationship between the original grid and its
dual: if the original grid has a nonempty probability q, then the dual grid has a
nonempty probability 1 q. For the existence of a crossing path, we have the
following lemma:
Lemma 12.8
As long as q > 23 , we have

P rq {Bwh } 1 (w + 1)eln(33q)h .
Proof. For the box Bwh and the corresponding dual box Swh , the comple-
is event S
ment of Bwh wh that there exists an open crossing path in Swh
between its top and bottom sides [80].
Note that the dual lattice has a nonempty probability q = 1 q < 13 . After
rotating the box by 90 , we can apply Lemma 12.7 to the dual lattice and obtain

P r{Swh } (w + 1)eln(3q)h = (w + 1)eln(33q)h .
This implies the result in this lemma.
We need the following definitions from [58]. An event A is an increasing

event if for any network where event A occurs, event A still occurs after adding
an edge in this network. Next we define the interior of A of depth r, Ir (A).
Event Ir (A) is an event such that for any network where Ir (A) occurs, event
A occurs after we change up to r arbitrary edges.
Now consider event A of having a crossing path within Bwh between its
left and right sides. It is easy to see that A is an increasing event. For this event,
Ir (A) is the event that (r + 1) edge-disjoint crossing paths exist. Theorem 2.45
in [58] asserts the following result:
For an increasing event A and any 0 q < q 1, we have
% &r
q
1 P rq {Ir (A)} (1 P rq {A}). (12.18)
q q
Based on the above results, we are now ready to give a proof of Theo-
rem 12.6.
Proof of Theorem 12.6. Denote Rm i as the event that there exists a crossing
path within the rectangle Rm between its left and right sides and let q = 2q 1.
i
Since q > 56 , we have 23 < q < q. By Lemma 12.8, we have

% / 1&
ln m(n)
P rq {Rm
i
} 1 (m(n) + 1) exp ln(3 3q) /( 2l(n))
m(n)
= 1 (m(n) + 1) exp (ln(6 6q) ln m(n)) . (12.19)
Thus, we have
P rq {Pmi ln m(n)} = 1 P rq {I ln m(n) (Rm

i
)}
% & ln m(n)
q
(1 P rq {Rm
i
)}
q q
% & ln m(n)
q
(m(n) + 1)eln(66q) ln m(n)
q q
q
=e
ln m(n) ln 1q +ln(m(n)+1)+ln(66q) ln m(n)
,
(12.20)
where the first inequality holds by (12.18), and the second inequality holds by
using ln m(n) ln m(n) and (12.19).
We now compute the probability of having at most ln m(n) edge-disjoint

i between its left and right sides. By (12.20),
crossing paths in any rectangle Rm
we have 5
P rq {Pm ln m(n)} = 1 P rq {Pm > ln m(n)}

m(n)
= 1 (P rq {Pmi > ln m(n)}) ln m(n)
m(n)
= 1 (1 P rq {Pmi ln m(n)}) ln m(n)
q m(n)
1(1e
ln m(n) ln 1q +ln(m(n)+1)+ln(66q) ln m(n)
) ln m(n) . (12.21)
To show that the above probability tends to zero as n , we need the

following result:

q
+ln(m(n)+1)+ln(66q) ln m(n)
> e m(n) .
ln m(n) ln 1q
1e (12.22)
To show this, note that ex 1 x for a small positive value x. We have

ln(1 e m(n) ) ln = ln ln m(n) > ln m(n), (12.23)
m(n)
where the last inequality holds since > 1 by (12.17). We also have
ln(m(n) + 1) > 2 ln m(n) (12.24)

ln(6 6q) ln m(n) > ln(6 6q) ln m(n). (12.25)
Thus, by (12.23), (12.24), and (12.25), we have

ln(1 e m(n) ) ln(m(n) + 1) ln(6 6q) ln m(n)
q
ln m(n) ln 1q
ln m(n) 2 ln m(n) ln(6 6q) ln m(n)
> q
ln m(n) ln 1q
ln(6 6q) 3
= q = .
ln 1q
Therefore, we have
q
ln m(n) ln + ln(m(n) + 1) + ln(6 6q) ln m(n) < ln(1 e m(n) ),
1q
which is equivalent to (12.22).
By (12.21) and (12.22), we have
m(n) 1
P rq {Pm ln m(n)} < 1 e m(n) ln m(n) = 1 e ln m(n) ,
which goes to zero as n .
5 Incidentally, we fix an error in [42] (Eq. (15)), where it was incorrectly stated that
m(n)
i ln m(n)}) ln m(n) .
P rq {Pm ln m(n)} = (P rq {Pm
Figure 12.7 The entry point for

the draining phase.
Highway entry point
Source
12.5.3 Deriving a feasible solution

Routing scheme In Section 12.5.2, we built a highway system for data trans-
missions in Phases (ii) and (iii),
where the hop length (the number of hops) for
each transmission is at most 2l(n) 2.
For Phases (i) and (iv), we now design a routing scheme and show that
the hop length is at most 2l(n) (d + 1), where d = ln m(n) + 1. We
adopt the following routing scheme for Phase (i) (the routing scheme for Phase
(iv) is similar). By Theorem 12.6, there are at least ln m(n) + 1 crossing
paths in each rectangle with height ln m(n)
m(n) almost surely as n , where
0
and are constants, and m(n) = 2 ln n
6 . Thus, we can slice each rect-
h
angle into horizontal strips of height m(n) (where h is an appropriate con-
stant) so that each strip corresponds to one crossing path. Note that a path
may not be fully contained in its corresponding strip. We can then impose that
the nodes in a strip communicate directly with the corresponding horizontal
path. Although a path may be outside its corresponding strip, it is never far-
ther than ln m(n)
m(n) from its corresponding strip. For each source in a strip, we
identify an entry point on the corresponding horizontal path such that the cell
that contains the entry point has the same horizontal position as the cell that
contains the source (see Fig. 12.7). Since each source finds its highway within
the same rectangle and at the same horizontal position, these two cells are at
most d = ln m(n)
m(n) /( 2l(n)) + 1 = ln m(n) + 1 (diagonal) cells away.
Thus, in all phases, transmissions between two cells is within at most d
(diagonal) cells away, where d = ln m(n) + 1 for Phases (i) and (iv) and
d = 1 for Phases (ii) and (iii).
Scheduling scheme We now design a time slot based scheduling. The idea
is that when a node transmits, other nodes that are within a certain distance
of this node cannot transmit simultaneously, whereas nodes that are outside
this distance can transmit. In particular, we use k 2 slots for scheduling, where
k = 2(d + 1). A set of cells that are allowed to transmit simultaneously in a
time slot is depicted in Fig. 12.8.
Figure 12.8 A set of cells that

can transmit simultaneously for
the case d = 1.
k=2(d+1)
=4
A lower bound for per-node throughput We can derive a lower bound for
per-node throughput by analyzing the achievable per-node throughput by the
above routing and scheduling schemes.
For the achievable rate at each hop, we have the following result:
Theorem 12.7
Denote cij (d) as the data rate at which node i in a cell can transmit to node
j in another cell that is at most d (diagonal) cells away. Then cij (d) is at
least (W ) almost surely as n , where d = 1 for Phases (ii) and (iii)
and d = ln m(n) + 1 for Phases (i) and (iv).
Proof. We focus on a particular cell. To analyze cij (d), we first find an upper
bound for the interference at the receiver j . We consider the transmitters in the
eight closest cells that are located at a Euclidean distance of at least l(n)(d +
1) from the receiver (see Fig. 12.8), along with the 16 next closest cells that
are located at a Euclidean distance of at least l(n)(3d + 3), and so on. By
extending the sum of the interferences to the entire two-dimensional area, we
have the following bound:

8iP 8P i
I (d) = .
(l(n)(2i 1)(d + 1)) (l(n)(d + 1)) (2i 1)
i=1 i=1
(12.26)
The sum in (12.26) converges because > 2. Thus, if d = 1, we have

I (d) ((l(n)d) ). (12.27)
Next, we find a lower bound for the signal received from transmitter
i. Since
the distance between a transmitter and its receiver is at most 2l(n)(d + 1),
the signal S(d) at the receiver satisfies
P
S(d) = ((l(n)d) ). (12.28)
( 2l(n)(d + 1))
S(d)
By combining (12.27) and (12.28), a lower bound on +I (d) is given by
S(d) ((l(n)d) )
= (1).
+ I (d) + ((l(n)d) )

S(d)
As a result, the achievable data rate cij (d) = W log2 1 + +I (d) has a con-
stant lower bound (W ) almost surely as n .
We can now analyze the achievable per-node throughput, which is bounded
by the bottleneck phases among Phases (i)(iv). We will show that the commu-
nication bottleneck resides in the highway Phases (ii) and (iii) with a per-node
throughput of ( 1n ).
For Phase (i), we have the following lemma:
Lemma 12.9
For transmissions
in Phase (i), each source node can achieve at least a rate
(ln n)3 almost surely as n .
W
Proof. To compute the rate at which nodes can communicate to the entry
points, we apply the second part of Theorem 12.7, which says that each node
can communicate to its entry point at least at a rate (W ). For transmissions in
Phase (i), where d = ln m(n) + 1, there are (2(d + 1))2 time slots required
for scheduling. By Lemma 12.5, the number of nodes in each cell is less than
ln m(n) almost surely as n . Therefore, the actual rate available
for each
node is at least (W )/((2(d + 1)) ln m(n)) = d 2 ln m(n) , i.e., the per-
2 W

node throughput is (lnWn)3 almost surely as n .
Since each node is the destination of exactly one source, we can prove the
following lemma by adopting the same approach as that in Lemma 12.9:
Lemma 12.10
For transmissions in Phase (iv), each destination
node can receive data from
the highway at least at a rate (ln n)3 almost surely as n .
W
Next, we consider the data rate along the highway. We first have the follow-
ing lemma:
Lemma 12.11
There are fewer than 2nh
m(n) nodes on each strip almost surely as n .
Proof. Let An be the event that there is at least one strip with no fewer
2nh
than m(n) nodes. Since the number of nodes in each strip is a Poisson ran-
nh
dom variable with parameter m(n) , by the union bounds and Lemma 12.6,
we have
# $ m(n)
2nh
1 nh
nh
e m(n) 1 m(n)
nh
e 2nh 1 e m(n)
nh
P r{An } e m(n)
m(n)
2nh
= e = ,
h m(n)
h 2 h 4
which goes to 0 as n .
The following lemma gives the achievable per-node throughput along the
highway:
Lemma 12.12
For transmissions in Phases (ii) and (iii), each node along the highway may
forward data for many source nodes. The achievable per-node throughput
for each source node is at least ( Wn ) almost surely as n .
Proof. We prove the result for the horizontal traffic in Phase (ii) (the proof for
Phase (iii) is similar and is thus omitted). For transmissions in Phase (ii), we
have d = 1 and the number of time slots required for scheduling is a constant
(2(d + 1))2 = 16. Further, the number of nodes selected for the highway in
each cell is one. Such a node can relay at most all the traffic generated in the
h
corresponding strip of height m(n) . According to Lemma 12.11, a node on a
horizontal highway may relay traffic for all nodes in this strip, which is at most
2nh
m(n) nodes. By applying the first part of Theorem 12.7, we conclude that one
source nodes data along the highway are at least ( W2nh ), i.e., the per-node
161 m(n)
rate is ( W
2nh ) = ( W m(n)
n ) = ( n ), almost surely as n .
W

m(n)
By Lemmas 12.9, 12.12, and 12.10, the per-node throughput is ( Wn )

almost surely as n . This proves Theorem 12.5.
313 12.7 Problems
In this chapter, we have described an asymptotic capacity analysis for wireless

ad hoc networks. Such an analysis addresses an achievable per-node through-
put when the number of nodes goes to infinity. We focused on so-called random
networks where each node is randomly deployed and each node has a randomly
chosen destination node. For such an asymptotic capacity analysis, the results
were derived in the form of (), O(), and (), where the underlying analy-
sis greatly differed from that in the previous chapters, which focused on opti-
mization problems for finite-sized networks. We showed that the asymptotic
capacity analysis heavily depends on the underlying interference model. In this
chapter, we considered three such models (i.e., the protocol model, the phys-
ical model, and the generalized physical model) and showed how to develop
asymptotic capacity bounds for each model.
12.7 Problems
12.1 Analyze the (), O(), and () relationships between the following
f (n) and g(n) functions:
(a) f (n) = 3n2 + 2n, g(n) = n2 ;
(b) f (n) = ln n, g(n) = n;
(c) f (n) = 1/n, g(n) = 1.
12.2 Describe the details for the protocol model, the physical model, and
the generalized physical model that we used in the asymptotic analysis in this
chapter.
12.3 For any two simultaneously active links i j and k l under the
protocol model, what is the maximum ratio between the interference from node
k to node j and the signal from node i to node j ?
12.4 In the proof of Theorem 12.1, we established that the network can
support O(r 2 (n)) transmissions at any time. Why is this not n (the number
of all source nodes)?
12.5 For a set of links whose receivers are inside a square of side-length
2
1/ r(n) , show that these links interfere with each other under the protocol
model.
12.6 If the cell size is changed to a(n) = 1/ lnnn in Lemma 12.1, prove that
there is still no empty cell almost surely as n . (Hint: You can prove this
by following a similar approach to that in the proof of Lemma 12.1.)
12.7 Display a possible route for the case that the S-D line crosses a gird
point in Fig. 12.9.
Figure 12.9 The straight line a(n)

connecting a source node s to
its destination node d.
Figure 12.10 A set of links in a

network. The solid lines
represent links and the dotted Link a
Link b
lines represent interference.
Link c
Link d
12.8 Under the protocol model, use a conflict graph to represent the interfer-
ence relationships between the links shown in Fig. 12.10, where the solid lines
represent links and the dotted lines represent interference. For example, there
exists a dotted line between the transmitter of link c and the receiver of link a,
which means that the receiver of link a is within the interference range of the
transmitter of link c . What is maximum node degree in the resulting conflict
graph?
12.9 What is the relationship between B and under the physical model?
12.10 Show that a capacity upper bound is O( Bn ) when all nodes use the
same transmit power under the physical model.
12.11 Suppose that all nodes use the same transmit power under the physical
model. Show that for any two successful simultaneous transmissions i j
315 12.7 Problems
and k l under the physical model, we have dkj (1 + )dij , where =

( 1/ 1).
12.12 Verify that f (x) = x (considered in the proof of Theorem 12.3) is a
convex function for x 0, where 2.
n
12.13 Prove that if f (x) is a convex function, then f
n n i=1 i xi
i=1 i f (xi ), where i=1 i = 1, i 0 for i = 1, . . . , n.
12.14 Compare the upper bound for a general case of independent power
control (in Theorem 12.3) and the upper bound for a special case of synchro-
nized power control (in Footnote 3). Which case has a tighter bound? Why can
we obtain a tighter bound for the other case?
12.15 For a random network generated by a Poisson point process with den-
sity n in a 1 1 area, show that if we divide the 1 1 area into small regions
of area lnnn , then every small region has at least one node almost surely as
n .
12.16 Why do we develop bounds for q in (12.14)? (Hint: Consider where
we have used q.)
12.17 Lemma 12.7 considers the case of q < 13 . Does the same result hold
for the case of q 13 ? If yes, why do we focus on the case of q < 13 in this
lemma? (Hint: Consider where we have used Lemma 12.7.)
References
1 E. Aarts and J. Korst, Simulated Annealing and Boltzman Machines, New York, NY,
John Wiley & Sons, 1989.
2 C. Ahn and R. Ramakrishna, A genetic algorithm for shortest path routing problem
and the sizing of populations, IEEE Transactions on Evolutionary Computation, vol. 6,
no. 6, pp. 566579, December 2002.
3 R.K. Ahuja, Algorithms for the minimax transportation problem, Naval Research
Logistics Quarterly, vol. 33, pp. 725739, 1986.
4 I.F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci, Wireless sensor net-
works: a survey, Computer Networks (Elsevier), vol. 38, no. 4, pp. 393422, 2002.
5 I. Akyildiz and X. Wang, A survey on wireless mesh networks, IEEE Communications
Magazine, vol. 43, no. 9, September 2005.
6 M. Alasti, K. Sayrafian-Pour, A. Ephremides, and N. Farvardin, Multiple description
coding in networks with congestion problem, IEEE Transactions on Information The-
ory, vol. 47, no. 3, pp. 891902, March 2001.
7 M. Alicherry, R. Bhatia, and L. Li, Joint channel assignment and
routing for throughput optimization in multi-radio wireless mesh net-
works, in Proc. ACM MobiCom, pp. 5872, Cologne, Germany, August
28September 2, August 2005.
8 J.G. Apostolopoulos, T. Wong, W. Tan, and S. Wee, On multiple description streaming
in content delivery networks, in Proc. IEEE INFOCOM, pp. 17361745, New York,
NY, June 2002.
9 T. Back, D. Fogel, and Z. Michalewicz (eds), Handbook of Evolutionary Computation,
New York, NY, Oxford University Press, 1997.
10 M.S. Bazaraa, J.J. Jarvis, and H.D. Sherali, Linear Programming and Network Flows,
4th edition, New York, NY, John Wiley & Sons Inc., 2010.
11 M.S. Bazaraa, H.D. Sherali, and C.M. Shetty, Nonlinear Programming: Theory and
Algorithms, 3rd edition, New York, John Wiley & Sons, 2006.
12 A.C. Begen, Y. Altunbasak, and O. Ergun, Multi-path selection for multiple descrip-
tion encoded video streaming, EURASIP Signal Processing: Image Communications,
vol. 20, no. 1, pp. 3960, January 2005.
13 A. Behzad and I. Rubin, Impact of power control on the performance of ad hoc wire-
less networks, in Proc. IEEE INFOCOM, pp. 102113, Miami, FL, March 1317,
2005.
317 References
14 D. Bertsekas and R. Gallager, Data Networks, Englewood Cliffs, NJ, Prentice Hall,
1992.
15 V. Bhandari and N.H. Vaidya, Capacity of multi-channel wireless networks with ran-
dom (c, f ) assignment, in Proc. ACM MobiHoc, pp. 229238, Montreal, QC, Canada,
September 914, 2007.
16 R. Bhatia and M. Kodialam, On power efficient communication over multi-hop wire-
less networks: joint routing, scheduling and power control, in Proc. IEEE INFOCOM,
pp. 14571466, Hong Kong, China, March 711, 2004.
17 E. Biglieri, R. Calderbank, A. Constantinides, A. Goldsmith, A. Paulraj, and H.V.
Poor, MIMO Wireless Communications, Cambridge, Cambridge University Press, Jan-
uary 2007.
18 A. Bletsas, A. Khisti, D. Reed, and A. Lippman, A simple cooperative diversity method
based on network path selection, IEEE Journal on Selected Areas in Communications,
vol. 24, no. 3, pp. 659672, March 2006.
19 C. Blum and A. Roli, Metaheuristics in combinatorial optimization: overview and con-
ceptual comparison, ACM Computing Surveys, vol. 35, no. 3, pp. 268308, Septem-
ber 2003.
20 T.X. Brown, H.N. Gabow, and Q. Zhang, Maximum flow-life curve for a wireless ad
hoc network, in Proc. ACM MobiHoc, pp. 128136, Long Beach, CA, October 45,
2001.
21 N. Bulusu and S. Jha (eds.), Wireless Sensor Networks: A Systems Perspective, Nor-
wood, MA, Artech House, 2005.
22 J. Cai, S. Shen, J.W. Mark, and A.S. Alfa, Semi-distributed user relaying algorithm for
amplify-and-forward wireless relay networks, IEEE Transactions on Wireless Commu-
nication, vol. 7, no. 4, pp. 13481357, April 2008.
23 J. Chakareski, S. Han, and B. Girod, Layered coding vs. multiple descrip-
tions for video streaming over multiple paths, in Proc. ACM Multimedia,
pp. 422431, Berkeley, CA, November 2003.
24 J.H. Chang and L. Tassiulas, Maximum lifetime routing in wireless sensor networks,
IEEE/ACM Transactions on Networking, vol. 12, issue 4, pp. 609619, August 2004.
25 P. Charnsethikul, The constrained minimax linear assignment problem, Optimization
Methods and Software, vol. 14, issues 1 & 2, pp. 3748, 2000.
26 C.C. Chen and D.S. Lee, A joint design of distributed QoS scheduling and power
control for wireless networks, in Proc. IEEE INFOCOM, Barcelona, Catalunya, Spain,
April 2329, 2006.
27 L. Chen, S.H. Low, M. Chiang, and J.C. Doyle, Cross-layer congestion control, rout-
ing and scheduling design in ad hoc wireless networks, in Proc. IEEE INFOCOM,
Barcelona, Catalunya, Spain, April 2329, 2006.
28 Concorde TSP Solver, http://www.tsp.gatech.edu/concorde/.
29 N. Clemens and C. Rose, Intelligent power allocation strategies in an unlicensed spec-
trum, in Proc. IEEE DySPAN, pp. 3742, Baltimore, MD, November 811, 2005.
30 T.J. Cormen, C.E. Leiserson, R.L. Rivest, and C. Stein, Introduction to Algorithms, 2nd
edition, The MIT Press, 2001.
31 IBM ILOG CPLEX Optimizer, http://www-01.ibm.com/software/integration/
optimization/cplex-optimizer/.
32 R.L. Cruz and A.V. Santhanam, Optimal routing, link scheduling and power control in
multi-hop wireless networks, in Proc. IEEE INFOCOM, pp. 702711, San Francisco,
CA, March 30April 3, 2003.
318 References
33 H. Dai, K. Ng, R.C. Wong, and M. Wu, On the capacity of multi-channel wireless net-
works using directional antennas, in Proc. IEEE INFOCOM, pp. 13011309, Phoenix,
AZ, April 1318, 2008.
34 I. Dietrich and F. Dressler, On the lifetime of wireless sensor networks, ACM Trans-
actions on Sensor Networks, vol. 5, no. 1, pp. 139, February 2009.
35 R. Draves, J. Padhye, and B. Zill, Routing in multi-radio, multi-hop wireless mesh
networks, in Proc. ACM MobiCom, pp. 114128, Philadelphia, PA, September 26
October 1, 2004.
36 J. Dreo, A. Petrowski, P. Siarry, and E. Taillard, Metaheuristics for Hard Optimization:
Methods and Case Studies, New York, Springer-Verlag, 2006.
37 T. Elbatt and A. Ephremides, Joint scheduling and power control for wireless ad-hoc
networks, in Proc. IEEE INFOCOM, pp. 976984, New York, NY, June 2327, 2002.
38 D. Eppstein, Finding the k shortest paths, SIAM Journal on Computing, vol. 28, no. 2,
pp. 652673, August 1999.
39 R. Etkin, A. Parekh, and D. Tse, Spectrum sharing for unlicensed bands, in
Proc. IEEE DySPAN, pp. 251258, Baltimore, MD, November 811, 2005.
40 FCC (Federal Communications Commission), Facilitating opportunities for flexible,
efficient, and reliable spectrum use employing cognitive radio technologies, notice of
proposed rule making and order, FCC 03-322.
41 G. J. Foschini and M. J. Gans, On limits of wireless communications in a fading envi-
ronment when using multiple antennas, Wireless Personal Communications, vol. 6,
pp. 311355, March 1998.
42 M. Franceschetti, O. Dousse, D.N.C. Tse, and P. Thiran, Closing the gap in the capac-
ity of wireless networks via percolation theory, IEEE Transaction on Information The-
ory, vol. 53, no. 3, pp. 10091018, March 2007.
43 D. Fritchman, A binary channel characterization using partitioned markov
chains, IEEE Transactions on Information Theory, vol. 13, no. 2,
pp. 221227, April 1967.
44 A.E. Gamal, J. Mammen, B. Prabhakar, and D. Shah, Throughput-delay trade-off
in wireless networks, in Proc. IEEE INFOCOM, pp. 464475, Hong Kong, China,
March 711, 2004.
45 G. Ganesan and Y.G. Li, Cooperative spectrum sensing in cognitive radio networks,
in Proc. IEEE DySPAN, pp. 137143, Baltimore, MD, November 811, 2005.
46 M.R. Garey and D.S. Johnson, Computers and Intractability: A Guide to the Theory of
NP-Completeness, New York, NY, W.H. Freeman and Company, 1979.
47 A. Gashemi and E. Sousa, Collaborative spectrum sensing for opportunistic access in
fading environments, in Proc. IEEE DySPAN, pp. 131136, Baltimore, MD, Novem-
ber 811, 2005.
48 D. Gesbert, M. Shafi, D. Shiu, P. J. Smith, and A. Naguib, From theory to practice: an
overview of MIMO space-time coded wireless systems, IEEE Journal on Select Areas
in Communications, vol. 21, no. 3, pp. 281302, April 2003.
49 E. Giler, Eric Giler demos wireless electricity, URL:http://www.ted.com/talks/eric
giler demos wireless electricity.html.
50 A. Giridhar and P.R. Kumar, Maximizing the functional lifetime of sensor networks,
in Proc. ACM/IEEE International Symposium on Information Processing in Sensor Net-
works, pp. 512, Los Angeles, CA, April 2005.
51 F. Glover and M. Laguna, Tabu Search. Boston, MA, Kluwer-Academic, 1997.
319 References
52 GNU Linear Programming Kit, http://www.gnu.org/software/glpk/.

53 N. Gogate, D. Chung, S.S. Panwar, and Y. Wang, Supporting image/video applica-
tions in a multihop radio environment using route diversity and multiple description
coding, IEEE Transactions on Circuits and Systems for Video Technology, vol. 12,
no. 9, pp. 777792, September 2002.
54 A.J. Goldsmith and S.-G. Chua, Adaptive coded modulation for fading channels,
IEEE Transactions on Communications, vol. 46, no. 5, pp. 595602, May 1998.
55 A. Goldsmith, S. A. Jafar, N. Jindal, and S. Vishwanath, Capacity limits of MIMO
channels, IEEE Journal on Select Areas in Communications, vol. 21, no. 1, pp. 684
702, June 2003.
56 A. Goldsmith, Wireless Communications, Cambridge, Cambridge University Press,
2005.
57 V. Goyal, Multiple description coding: Compression meets the network, IEEE Signal
Processing Magazine, vol. 18, pp. 7493, September 2001.
58 G. Grimmett, Percolation, 2nd edition, New York, Springer-Verlag, 1999.
59 P. Gupta and P.R. Kumar, Critical power for asymptotic connectivity in wireless net-
works, in Stochastic Analysis, Control, Optimization and Applications: A Volume in
Honor of W.H. Fleming, W.M. McEneany, G. Yin, and Q. Zhang (eds.), Boston, MA,
Birkhauser, pp. 547566, 1998.
60 P. Gupta and P.R. Kumar, The capacity of wireless networks, IEEE Transactions on
Information Theory, vol. 46, no. 2, pp. 388404, March 2000.
61 A. Keshavarz-Haddad, V. Ribeiro, and R. Riedi, Broadcast capacity in multihop wire-
less networks, in Proc. ACM MobiCom, pp. 239250, Los Angeles, CA, Septem-
ber 2326, 2006.
62 A. Keshavarz-Haddad and R. Riedi, On the broadcast capacity of multihop wire-
less networks: interplay of power, density and interference, in Proc. IEEE SECON,
pp. 314323, San Diego, CA, June 1821, 2007.
63 A. Keshavarz-Haddad and R. Riedi, Multicast capacity of large homogeneous multi-
hop wireless networks, in Proc. International Symposium on Modeling and Optimiza-
tion in Mobile, Ad Hoc, and Wireless Networks (WiOpt), pp. 116124, Berlin, Germany,
April 2008.
64 S. Haykin, Adaptive Filter Theory, Englewood Cliffs, NJ, Prentice-Hall, 1996.
65 S. Haykin, Cognitive radio: brain-empowered wireless communications, IEEE Jour-
nal on Selected Areas in Communications, vol. 23, no. 2, pp. 201220, February 2005.
66 S. He, J. Chen, F. Jiang, D.K.Y. Yau, G. Xing, and Y. Sun, Energy provisioning in
wireless rechargeable sensor networks, in Proc. IEEE INFOCOM, pp. 20062014,
Shanghai, China, April 2011.
67 W. Heinzelman, Application-specific protocol architectures for wireless networks,
Ph.D. thesis, Department of Electrical Engineering and Computer Science, Mas-
sachusetts Institute of Technology, June 2000.
68 J.-B. Hiriart-Urruty and C. Lemarechal, Fundamentals of Convex Analysis, Berlin,
Springer-Verlag, 2001.
69 Y.T. Hou, Y. Shi, and H.D. Sherali, On node lifetime problem for energy-constrained
wireless sensor networks, ACM/Springer Mobile Networks and Applications, vol. 10,
no. 6, pp. 865878, December 2005.
70 Y.T. Hou and Y. Shi, Variable bit rate flow routing in wireless sensor networks, IEEE
Transactions on Wireless Communications, vol. 56, no. 6, pp. 21402148, June 2007.
320 References
71 Y.T. Hou, Y. Shi, and H.D. Sherali, Rate allocation and network lifetime problems
for wireless sensor networks, IEEE/ACM Transactions on Networking, vol. 16, no. 2,
pp. 321334, April 2008.
72 Y.-C. Hu and D. Johnson, Design and demonstration of live audio and
video over multihop wireless ad hoc networks, in Proc. IEEE MILCOM, pp.
710, Anaheim, CA, October 2002.
73 J. Huang, R.A. Berry, and M.L. Honig, Spectrum sharing with distributed interference
compensation, in Proc. IEEE DySPAN, pp. 8893, Baltimore, MD, November 811,
2005.
74 O. Ileri, D. Samardzija, T. Sizer, and N.B. Mandayam, Demand responsive pricing
and competitive spectrum allocation via a spectrum server, in Proc. IEEE DySPAN,
pp. 194202, Baltimore, MD, November 811, 2005.
75 S.-W. Jeon, N. Devroye, M. Vu, S.-Y. Chung, and V. Tarokh, Cognitive networks
achieve throughput scaling of a homogeneous network, in Proc. ICST International
Symposium on Modeling and Optimization in Mobile, Ad Hoc, and Wireless Networks
(WiOpt), Seoul, Korea, June 2627, 2009.
76 X. Jiang, J. Polastre, and D. Culler, Perpetual environmentally powered sensor net-
works, in Proc. ACM/IEEE International Symposium on Information Processing in
Sensor Networks, pp. 463468, Los Angeles, CA, April 2005.
77 A. Kansal, J. Hsu, S. Zahedi, and M.B. Srivastava, Power management in energy har-
vesting sensor networks, ACM Transactions Embedded Computing Systems, vol. 6,
no. 4, article 32, September 2007.
78 K. Kar, M. Kodialam, T.V. Lakshman, and L. Tassiulas, Routing for network capac-
ity maximization in energy-constrained ad-hoc networks, in Proc. IEEE INFOCOM,
pp. 673681, March 30April 3, 2003, San Francisco, CA.
79 F. Kelly, A. Maulloo, and D. Tan, Rate control for communication networks: shadow
prices, proportional fairness and stability, in Journal of the Operational Research Soci-
ety, vol. 49, no. 3, pp. 237252, March 1998.
80 H. Kesten, Percolation Theory for Mathematicians, Boston, MA, Birkhauser, 1982.
81 J. Kleinberg and E. Tardos, Algorithm Design, Boston, MA, Addison-Wesley Longman
Publishing, 2005.
82 M. Kodialam and T. Nandagopal, Characterizing the capacity region in multi-radio
multi-channel wireless mesh networks, in Proc. ACM MobiCom, pp. 7387, Cologne,
Germany, August 28September 2, 2005.
83 M. Kodialam and T. Nandagopal, Characterizing achievable rates in multi-hop wire-
less mesh networks with orthogonal channels, IEEE/ACM Transactions on Network-
ing, vol. 13, no. 4, pp. 868880, August 2005.
84 G. Kramer, I. Maric, and R.D. Yates, Cooperative communications, Foundations and
Trends in Networking, Now Publishers, June 2007.
85 A. Kurs, A. Karalis, R. Moffatt, J.D. Joannopoulos, P. Fisher, and M. Soljacic, Wireless
power transfer via strongly coupled magnetic resonances, Science, vol. 317, no. 5834,
pp. 8386, 2007.
86 P. Kyasanur and N.H. Vaidya, Capacity of multi-channel wireless networks: impact
of number of channels and interfaces, in Proc. ACM MobiCom, pp. 4357, Cologne,
Germany, August 28September 2, 2005.
87 J.N. Laneman, D.N.C. Tse, and G.W. Wornell, Cooperative diversity in wireless net-
works: efficient protocols and outage behavior, IEEE Transactions on Information
Theory, vol. 50, no. 12, pp. 30623080, December 2004.
321 References
88 L. Lasdon, Optimization Theory for Large Scale Systems, New York, NY, Macmillan.,
1970.
89 E. Lawler, Combinatorial Optimization: Networks and Matroids, Mineola, NY, Dover
Publications Inc., 2001.
90 J.-W. Lee, M. Chiang, and R.A. Calderbank, Price-based distributed algorithm for
optimal rate-reliability tradeoff in network utility maximization, IEEE Journal on
Selected Areas in Communications, vol. 24, issue 5, pp. 962976, May 2006.
91 X.-Y. Li, J.-Z. Zhao, Y.-W. Wu, S.-J. Tang, X.-H. Xu, and X.F. Mao, Broadcast capac-
ity for wireless ad hoc networks, in Proc. IEEE MASS, pp. 114123, Atlanta, GA,
September 29October 2, 2008.
92 X.-Y. Li, Multicast capacity of wireless ad hoc networks, IEEE/ACM Transactions
Networking, vol. 17, no. 3, pp. 950961, June 2009.
93 Z. Li, Y. Peng, W. Zhang, and D. Qiao, J-RoC: A joint routing and charging scheme
to prolong sensor network lifetime, in Proc. IEEE ICNP, pp. 373382, Vancouver,
Canada, October 1720, 2011.
94 D. Linden and T.B. Reddy (eds.), Handbook of Batteries, 3rd edition, McGraw-Hill,
2002.
95 LINDO SystemsOptimization Software, http://www.lindo.com/.
96 X. Liu and W. Wang, On the characteristics of spectrum-agile communication net-
works, in Proc. IEEE DySPAN, pp. 214223, Baltimore, MD, November 811, 2005.
97 J. Liu, D. Goeckel, and D. Towsley, Bounds on the gain of network coding and broad-
casting in wireless networks, in Proc. IEEE INFOCOM, pp. 724732, Anchorage, AK,
May 612, 2007.
98 S.H. Low, A duality model of TCP and queue management algorithms, IEEE/ACM
Transactions on Networking, vol. 11, issue 4, pp. 525536, August 2003.
99 H. Luss and D.R. Smith, Resource allocation among competing activities: a lexico-
graphic minimax approach, Operations Research Letters, vol. 5, no. 5, pp. 227231,
November 1986.
100 J. R. Magnus and H. Neudecker, Matrix Differential Calculus with Applications in
Statistics and Economics, New York, NY, Wiley, 1999.
101 N. Malpani and J. Chen, A note on practical construction of maximum bandwidth
paths, Information Processing Letters, vol. 83, pp. 175180, August 2002.
102 S. Mao, S. Lin, S.S. Panwar, Y. Wang, and E. Celebi, Video transport over ad hoc
networks: multistream coding with multipath transport, IEEE Journal on Select Areas
in Communications, vol. 21, no. 10, pp. 17211737, December 2003.
103 S. Mao, S. Kompella, Y.T. Hou, H.D. Sherali, and S.F. Midkiff, Routing for concur-
rent video sessions in ad hoc networks, IEEE Transactions on Vehicular Technology,
vol. 55, issue 1, pp. 317327, January 2006.
104 S. Mao, Y.T. Hou, X. Cheng, H.D. Sherali, S.F. Midkiff, and Y.-Q. Zhang, On routing
for multiple description video over wireless ad hoc networks, IEEE Transactions on
Multimedia, vol. 8, issue 5, pp. 10631074, October 2006.
105 I. Maric and R.D. Yates, Cooperative multihop broadcast for wireless networks,
IEEE Journal on Selected Areas in Communications, vol. 22, issue 6, pp. 10801088,
August 2004.
106 M. McHenry and D. McCloskey, Spectrum occupancy measurements: Republican
National Convention, New York, NY, August 30September 3, 2004, available at
http://www.sharedspectrum.com/wp-content/uploads/4 NSF NYC Report.pdf.
322 References
107 R. Meester and R. Roy, Continuum Percolation, Cambridge, Cambridge University

Press, 1996.
108 N. Megiddo, Linear-time almorithm for linear programming in r 3 and related prob-
lems, SIAM Journal on Computing, vol. 12, pp. 759776, 1983.
109 E.C. van der Meulen and P. Vanroose. The capacity of a relay channel, both with and
without delay, IEEE Transactions on Information Theory, vol. 53, no. 10, pp. 3774
3776, October 2007.
110 Z. Michalewicz and D.B. Fogel, How to Solve It: Modern Heuristics, 2nd edition, New
York, Springer-Verlag, 2004.
111 S.M. Mishra, A. Sahai, and R.W. Brodersen, Cooperative sensing among cognitive
radios, in Proc. IEEE International Conference on Communications, pp. 16581663,
Istanbul, Turkey, June 1115, 2006.
112 J. Mitola III, Cognitive radio: an integrated agent architecture for software defined
radio, Ph.D. thesis, KTH Royal Institute of Technology, 2000.
113 G.L. Nemhauser and L.A. Wolsey, Integer and Combinatorial Optimization, New York,
NY, John Wiley & Sons, 1999.
114 A. Nemirovski, Interior Point Polynomial Time Methods in Convex Programming,
2004, available at http://www2.isye.gatech.edu/nemirovs/Lect IPM.pdf.
115 N. Nie and C. Comaniciu, Adaptive channel allocation spectrum etiquette for cognitive
radio networks, in Proc. IEEE DySPAN, pp. 269278, Baltimore, MD, November 8
11, 2005.
116 R. Ogier, F. Templin, and M. Lewis, Topology dissemination based on reverse-path
forwarding (TBRPF), February 2004, IETF RFC 3684.
117 L. Ozarow, On a source coding problem with two channels and three receivers, Bell
System Technical Journal, vol. 59, no. 10, pp. 8491, December 1980.
118 A. Ozgur, O. Leveque, and D.N.C. Tse, Hierarchical cooperation achieves optimal
capacity scaling in ad hoc networks, IEEE Transaction on Information Theory, vol. 53,
no. 10, pp. 35493572, October 2007.
119 M. Padberg and G. Rinaldi, A branch-and-cut algorithm for the resolution of large-
scale symmetric traveling salesman problems, SIAM Review, vol. 33, no. 1, pp. 60
100, 1991.
120 P. Papadimitratos, Z. Haas, and E. Sirer, Path set selection in mobile ad hoc networks,
in Proc. ACM MobiHoc, pp. 111, Lausanne, Switzerland, June 2002.
121 G. Park, T. Rosing, M.D. Todd, C.R. Farrar, and W. Hodgkiss, Energy harvesting
for structural health monitoring sensor networks, Journal of Infrastructure Systems,
vol. 14, no. 1, pp. 6479, March 2008.
122 A.J. Paulraj, D.A. Gore, R.U. Nabar, and H. Bolcskei, An overview of MIMO com-
munications a key to gigabit wireless, Proceedings of the IEEE, vol. 92, no. 2, pp.
198218, February 2004.
123 C. Peng, H. Zheng, and B.Y. Zhao, Utilization and fairness in spectrum assignment
for opportunistic spectrum access, ACM/Springer Mobile Networks and Applications,
vol. 11, issue 4, pp. 555576, August 2006.
124 Y. Peng, Z. Li, W. Zhang, and D. Qiao, Prolonging sensor network lifetime through
wireless charging, in Proc. IEEE RTSS, pp. 129139, San Diego, CA, November 30
December 3, 2010.
125 PowerCast, http://www.powercastco.com.
323 References
126 A.P. Punnen and K.P.K. Nair, Improved complexity bound for the maximum cardi-
nality bottleneck bipartite matching problem, Elsevier Discrete Applied Mathematics,
vol. 44, pp. 9193, 1994.
127 R. Ramanathan and R. Rosales-Hain, Topology control of multihop wireless networks
using transmit power adjustment, in Proc. IEEE INFOCOM, pp. 404413, 2000, Tel
Aviv, Israel, March 2630.
128 A. Raniwala and T. Chiueh, Architecture and algorithms for an IEEE 802.11-based
multi-channel wireless mesh network, in Proc. IEEE INFOCOM, pp. 22232234,
Miami, FL, March 1317, 2005.
129 T.S. Rappaport, Wireless Communications: Principles and Practice, Upper Saddle
River, NJ, Prentice Hall, 1996.
130 J.H. Reed, Software Radio: A Modern Approach to Radio Engineering, Upper Saddle
River, NJ, Prentice Hall, May 2002.
131 V. Rodoplu and T.H. Meng, Minimum energy mobile wireless networks, IEEE Jour-
nal on Selected Areas in Communications, vol. 17, no. 8, pp. 13331344, August 1999.
132 J. B. Rosen, The gradient projection method for nonlinear programming, part I: linear
constraints, SIAM Journal on Applied Mathematics, vol. 8, pp. 181217, 1960.
133 A. Sankar and Z. Liu, Maximum lifetime routing in wireless ad-hoc networks, in
Proc. IEEE INFOCOM, pp. 10891097, Hong Kong, China, March 711, 2004.
134 Wireless Innovation Forum, http://www.sdrforum.org.
135 E. Setton, Y. Liang, and B. Girod, Adaptive multiple description video streaming over
multiple channels with active probing, in Proc. IEEE ICME, Baltimore, MD, July
2003.
136 A. Sendonaris, E. Erkip, and B. Aazhang, User cooperation diversity part I: system
description, IEEE Transactions on Communications, vol. 51, no. 11, pp. 19271938,
November 2003.
137 A. Sendonaris, E. Erkip, and B. Aazhang, User cooperation diversity part II: imple-
mentation aspects and performance analysis, IEEE Transactions on Communications,
vol. 51, no. 11, pp. 19391948, November 2003.
138 S. Shakkottai, X. Liu, and R. Srikant, The multicast capacity of large multihop wireless
networks, in Proc. ACM MobiHoc, pp. 247255, Montreal, Quebec, Canada, Septem-
ber 914, 2007.
139 S. Sharma, Y. Shi, Y.T. Hou, and S. Kompella, An optimal algorithm for relay node
assignment in cooperative ad hoc networks, IEEE/ACM Transactions on Networking,
vol. 19, issue 3, pp. 879892, June 2011.
140 H.D. Sherali and W.P. Adams, A hierarchy of relaxations between the continuous and
convex hull representations for zero-one programming problems, SIAM Journal on
Discrete Mathematics, vol. 3, no. 3, pp. 411430, 1990.
141 H.D. Sherali and C.H. Tuncbilek, A global optimization algorithm for polynomial pro-
gramming problems using a reformulation-linearization technique, Journal of Global
Optimization, vol. 2, pp. 101112, 1992.
142 H.D. Sherali and W.P. Adams, A hierarchy of relaxations and convex hull characteri-
zations for mixed-integer zero-one programming problems, Discrete Applied Mathe-
matics, vol. 52, pp. 83106, 1994.
143 H.D. Sherali and C.H. Tuncbilek, Comparison of two reformulation-linearization tech-
nique based linear programming relaxations for polynomial programming problems,
Journal of Global Optimization, vol. 10, pp. 381390, 1997.
324 References
144 H.D. Sherali, Global optimization of nonconvex polynomial programming problems

having rational exponents, Journal of Global Optimization, vol. 12, no. 3, pp. 267283,
1998.
145 H. Sherali, K. Ozbay, and S. Subramanian, The time-dependent shortest pair of disjoint
paths problem: complexity, models, and algorithms, Networks, vol. 31, no. 4, pp. 259
272, December 1998.
146 H.D. Sherali and W.P. Adams, A Reformulation-Linearization Technique for Solv-
ing Discrete and Continuous Nonconvex Problems, Dordrecht/Boston/London, Kluwer
Academic Publishers, 1999.
147 H.D. Sherali, On mixed-integer zero-one representations for separable lower-
semicontinuous piecewise-linear functions, Operations Research Letters, vol. 28,
no. 4, pp. 155160, 2001.
148 H.D. Sherali and L. Liberti, Reformulation-linearization technique for global opti-
mization, in Encylopedia of Optimization, 2nd edition, P. Pardalos and C. Floudas
(eds.) Springer, Berlin, pp. 32633268, 2008.
149 H.D. Sherali, Reformulation-linearization technique for MIPs, in Wiley Encyclope-
dia of Operations Research and Management Science, Invited Advanced Topic paper,
February 2011.
150 Y. Shi and Y.T. Hou, Optimal power control for multi-hop software defined radio net-
works, in Proc. IEEE INFOCOM, pp. 16941702, Anchorage, AL, May 612, 2007.
151 Y. Shi and Y.T. Hou, Theoretical results on base station movement problem for sensor
network, in Proc. IEEE INFOCOM, pp. 376384, Phoenix, AZ, April 1417, 2008.
152 Y. Shi, Y.T. Hou, J. Liu, and S. Kompella, Bridging the gap between protocol and phys-
ical models for wireless networks, IEEE Transactions on Mobile Computing, 2012.
153 K. Sohrabi, J. Gao, V. Ailawadhi, and G. Pottie, Protocols for self-organizing of a
wireless sensor network, IEEE Personal Communications Magazine, vol. 7, pp. 16
27, October 2000.
154 M.E. Steenstrup, Opportunistic use of radio-frequency spectrum: a network perspec-
tive, in Proc. IEEE DySPAN, pp. 638641, Baltimore, MD, November 811, 2005.
155 I. E. Telatar, Capacity of multi-antenna Gaussian channels, European Transactions
on Telecommunications, vol. 10, no. 6, pp. 585596, November 1999.
156 N. Tesla, Apparatus for transmitting electrical energy, US patent number 1,119,732,
issued in December 1914.
157 B. Tong, Z. Li, G. Wang, and W. Zhang, How wireless power charging technology
affects sensor network deployment and routing, in Proc. IEEE ICDCS, pp. 438447,
Genoa, Italy, June 2010.
158 D. Ugarte and A.B. McDonald, On the capacity of dynamic spectrum access enabled
networks, in Proc. IEEE DySPAN, pp. 630633, Baltimore, MD, November 811,
2005.
159 T.A. Vanderelli, J.G. Shearer, and J.R. Shearer, Method and apparatus for a wireless
power supply, US patent number 7,027,311, issued in April 2006.
160 E. Visostky, S. Kuffner, and R. Peterson, On collaborative detection of TV transmis-
sions in support of dynamic spectrum sharing, in Proc. IEEE DySPAN, pp. 338345,
Baltimore, MD, November 811, 2005.
161 A. Volgenant and C.W. Duin, Improved polynomial algorithms for robust bottleneck
problems with interval data, Elsevier Computers and Operations Research, vol. 37,
issue 5, pp. 909915, May, 2010.
325 References
162 M. Vu, N. Devroye, M. Sharif, and V. Tarokh, Scaling laws of cognitive networks, in
Proc. ICST International Conference on Cognitive Radio Oriented Wireless Networks
and Communications (CrownCom), pp. 28, Orlando, FL, July 31August 3, 2007.
163 W. Wang, V. Srinivasan, and K.C. Chua, Using mobile relays to prolong the lifetime of
wireless sensor networks, in Proc. ACM MobiCom, Cologne, pp. 270283, Germany,
August 2005.
164 W. Wang, V. Srinivasan, and K.C. Chua, Extending the lifetime of wireless sensor net-
works through mobile relays, IEEE/ACM Transactions on Networking, vol. 16, no. 5,
pp. 11081120, October 2008.
165 Z. Wang, H.R. Sajadpour, and J.J. Garcia-Luna-Aceves, The capacity and energy effi-
ciency of wireless ad hoc networks with multi-packet reception, in Proc. ACM Mobi-
Hoc, pp. 179188, Hong Kong, China, May 2630, 2008.
166 R. Wattenhofer, L. Li, P. Bahl, and Y.-M. Wang, Distributed topology control for power
efficient operation in multihop wireless ad hoc networks, in Proc. IEEE INFOCOM,
pp. 13881397, Anchorage, AK, April 2226, 2001.
167 E. Welzl, Smallest enclosing disks, Lecture Notes in Computer Science (LNCS),
vol. 555, pp. 359370, 1991.
168 D.B. West, Introduction to Graph Theory, Upper Saddle River, NJ, Prentice Hall, 2001.
169 J. Winters, On the capacity of radio communication systems with diversity in a
Rayleigh fading environment, IEEE Journal on Selected Areas in Communications,
vol. 5, no. 5, pp. 871878, June 1987.
170 J. H. Winters, Smart antenna techniques and their application to wireless ad hoc
networks, IEEE Wireless Communications Magazine, vol. 13, no. 4, pp. 7783,
August 2006.
171 WiTricity Corp., http://www.witricity.com.
172 Wireless Power Consortium, http://www.wirelesspowerconsortium.com/.
173 A. Wyglinski, M. Nekovee, and Y.T. Hou (eds.), Cognitive Radio Communications and
Networks: Principles and Practice, Elsevier, December 2009.
174 C. Xin, B. Xie, and C.-C. Shen, A novel layered graph model for topology formation
and routing in dynamic spectrum access networks, in Proc. IEEE DySPAN, pp. 308
317, Baltimore, MD, November 811, 2005.
175 S. Ye and R. S. Blum, Optimized signaling for MIMO interference systems with
feedback, IEEE Transactions on Signal Processing, vol. 51, no. 11, pp. 28392848,
November 2003.
176 C. Yin, L. Gao, and S. Cui, Scaling laws of overlaid wireless networks: a cognitive
radio network vs. a primary network, IEEE/ACM Transactions on Networking, vol. 18,
no. 4, pp. 13171329, August 2010.
177 H. Zhang and J.C. Hou, Capacity of wireless ad-hoc networks under ultra wide band
with power constraint, in Proc. IEEE INFOCOM, pp. 455465, Miami, FL, March 13
17, 2005.
178 F. Zhang, X. Liu, S.A. Hackworth, R.J. Sclabassi, and M. Sun, In vitro and in
vivo studies on wireless powering of medical sensors and implantable devices, in
Proc. IEEE/NIH Life Science Systems and Applications Workshop (LiSSA), pp. 8487,
Bethesda, MD, April 2009.
179 J. Zhao, H. Zheng, and G. Yang, Distributed coordination in dynamic spectrum alloca-
tion networks, in Proc. IEEE DySPAN, pp. 259268, Baltimore, MD, November 811,
2005.
326 References
180 Y. Zhao, R.S. Adve, and T.J. Lim, Improving amplify-and-forward relay networks:
optimal power allocation versus selection, in Proc. IEEE International Symposium on
Information Theory, pp. 12341238, Seattle, WA, July 914, 2006.
181 R. Zheng, Information dissemination in power-constrained wireless network, in
Proc. IEEE INFOCOM, Barcelona, Catalunya, Spain, April 2329, 2006.
Index
(1 + )-optimal, 192 partitioning variable, 4, 96, 119

(1 + )-optimal approximation algorithm, 192 branching step, 96
(1 )-optimal, 5, 192, 233 dichotomous branching, 96
(1 )-optimal approximation algorithm, 5, broadcast capacity, 284
192, 198, 240
(1 )-optimal solution, 5, 39, 95, 125, 191, capacity maximization problem, 125
212 capacity scaling law, 287, 295, 300
(1 )-optimal algorithm, 225 capacity scaling law analysis, 283
(1 )-optimality, 212, 225 cell-based approach, 289
cell-based routing, 289
affine function, 38 channel gain, 296
aggregate capacity distance, 296 channel gain matrix, 42
aggregate rate (AR), 288 cognitive radio (CR), 2, 100
algorithms complexity, 61 cognitive radio network (CRN), 100, 244, 284
almost surely, 285 complexity, 61
amplify-and-forward (AF), 63 algorithms complexity, 61
approximation algorithm, 5, 191 problems complexity, 61
(1 + )-optimal approximation algorithm, concave function, 38
192 conflict graph, 295, 299
(1 )-optimal approximation algorithm, conflict relationship, 287
5, 192 conjugate (sub)gradient method, 39
constant-factor approximation algorithm, conjugate transpose, 42, 64
191 constant-factor approximation algorithm, 191
asymptotic capacity, 7 constrained mobile base station (C-MB)
asymptotic capacity analysis, 7, 283, 285, 313 problem, 212, 223
convex function, 38
bandwidth-footprint-product (BFP), 108, 250 convex hull relaxation, 111
base station, 5, 13, 155, 192, 211 three tangential supports, 133
base station placement problem, 5, 193, 208 convex optimization, 40
basic local search, 7, 260 dual decomposition, 46
block diagonal matrix, 42 dual problem, 4, 47
bound-factor constraint, 123, 124, 133 convex programming, 3, 38, 39, 46, 96
bounding step, 96 strong duality, 40
branch-and-bound, 4, 95, 132 weak duality, 40
bounding step, 96 convex set, 38
branching step, 96 cooperative communications (CC), 2, 4, 61, 62
fathoming a subproblem, 99 amplify-and-forward (AF), 63
local search, 4, 96, 113, 135 decode-and-forward (DF), 64
partitioning problem, 96 cooperative relaying, 62
327
328 Index
core optimization space, 5, 125, 134, 146 generalized physical model, 7, 285
core optimization variable, 125 physical model, 7, 285
cross-layer convex optimization, 3 protocol model, 7, 285
cross-layer optimization, 3, 45, 119, 129, 263 signal-to-interference-and-noise ratio
crossover, 262, 271 (SINR) model, 5, 125
cutting-plane method, 4, 41, 49 interference range, 103, 247, 254, 285, 287
interior, 307
decision variable, 11, 247, 251 interior point method, 12, 39
decode-and-forward (DF), 64 iterative improvement, 260
decomposition, 4, 41
deflected (sub)gradient method, 39 Lagrangian dual problem, 39, 47, 55
determinant function, 64 Lagrangian dual variable, 47
dichotomous branching, 96 Lagrangian function, 40
double description (DD) video, 263 Lagrangian multiplier, 39, 47
dual decomposition, 46 lexicographic max-min (LMM), 3, 14, 19
dual grid, 306 linear approximation, 5, 148, 151, 179
dual problem, 4, 12, 47 grid linearization, 148
dual simplex method, 12 piecewise linear approximation, 150, 152,
dual variable, 39 167
duality relationship, 12 linear marking, 72, 75
linear programming (LP), 3, 11, 13
empty cell, 290, 304 dual problem, 12
energy-harvesting, 151 dual simplex method, 12
entry point, 309
parametric analysis (PA), 14, 22
expectation function, 64
standard form, 12
linear relaxation, 5, 48, 111, 122, 152
fathoming a subproblem, 99
convex hull relaxation, 111
fictitious cost point (FCP), 6, 193, 201
LMM node lifetime, 14, 25
fitness function, 261, 270
mirror relationship, 14, 27
flow routing, 3, 14, 101, 125, 152, 194, 212,
LMM rate allocation, 14, 19
244
mirror relationship, 14, 27
generalized physical model, 7, 284, 285, 300 LMM-optimal solution, 19
genetic algorithm (GA), 7, 261 local optimal solution, 39
crossover, 262, 271 local search, 4, 96, 113, 135
fitness function, 261 basic local search, 7, 260
mutation, 262, 272 location-dependent problem, 6, 212, 215, 223,
Tournament selection, 271 241
global optimal solution, 39, 99 LP duality theory, 51
gradient method, 39
conjugate (sub)gradient method, 39 magnetic resonant coupling, 151
deflected (sub)gradient method, 39 max-min rate allocation, 14, 19
gradient projection method, 39 metaheuristic, 6, 260, 263, 277
gradient projection method, 39 genetic algorithm (GA), 7, 261
grid linearization, 148 iterative improvement, 260
simulated annealing (SA), 7 , 260, 263
Hamiltonian cycle, 152, 162, 174 tabu search (TS), 7, 261, 263
hidden terminal, 248 MIMO link, 43
highway path, 301, 302 MIMO network, 4, 40, 44, 57
highway system, 301, 302, 304, 309 mirror relationship, 14, 27
mixed-integer linear programming (MILP), 4,
identity matrix, 42, 64 13, 39, 68, 243
increasing event, 307 mixed-integer nonlinear programming
inner product, 42 (MINLP), 5, 110, 125, 251
integer linear program (ILP), 13 mixed-integer optimization problem, 243, 256
interference model, 7, 283, 285, 301, 313 mobile base station problem, 5, 212
329 Index
constrained mobile base station (C-MB) Percolation theory, 300, 306

problem, 212, 223 dual grid, 306
location-dependent problem, 6 physical model, 7, 125, 284, 285, 295
time-dependent problem, 6 piecewise linear approximation, 150, 152, 167
unconstrained mobile base station (U-MB) piecewise linear function, 149, 150
problem, 212, 223 positive semidefinite, 42
monomial, 5, 122 power control, 3, 101, 125, 194, 295
multi-channel, 284 synchronized power control, 284
multi-channel multi-radio (MC-MR), 284 power control problem, 100, 101, 119
multi-channel single-radio (MC-SR), 284 power spectral density, 247
multi-path routing, 7, 128, 213, 275, 277 primal-dual algorithm, 12
multicast capacity, 284 problem formulation, 2, 46, 109, 130, 215,
multiple description (MD) coding, 262 251, 263
double description (DD) video, 263 problem space, 38
multiple-input multiple-output (MIMO), 2, 40, problems complexity, 61
284 protocol model, 7, 284, 285, 287
multiple-packet reception (MPR), 284
mutation, 262, 272 random network, 7, 283, 301, 313
reformulation, 2, 109, 129, 166
near-optimal solution, 1, 7, 148, 174, 234, 244, Reformulation-Linearization Technique
277 (RLT), 5, 122, 144
network capacity, 13
bound-factor constraint, 123, 124, 133
network coding, 284
relaxation, 96, 251
network lifetime, 13, 154, 194, 213
linear relaxation, 5, 48
network throughput, 40
relaxation error, 96, 114, 137
network utility, 41
relay node assignment problem, 61, 65, 67, 89
node-link incidence matrix (NLIM), 44
renewable energy cycle, 152, 156, 158
noise density, 108, 128, 250
right-hand-side (RHS), 12, 45
nonconvex optimization, 4, 119, 208
routing, 289
nonconvex programming, 122, 196
nonlinear optimization, 3, 57, 148, 152, 179
scheduling, 3, 101, 125, 244, 289, 293, 309
nonlinear programming (NLP), 5, 193
self-interference, 41, 106
nonpolynomial program, 215
sensor network, 194
NP-complete, 95
separable programming, 148
NP-hard, 4, 68, 95, 131, 152, 193, 244, 269
sequential fixing (SF), 6, 243
objective function, 7, 11, 39, 66, 148, 196, sequential quadrification process, 123
213, 245, 268 Shannon capacity, 40
open path, 304 Shannon capacity formula, 125, 285
optimal solution, 1, 22, 58, 61, 99, 131, 152, signal-to-interference-and-noise ratio (SINR),
191, 225, 262 285
(1 )-optimal solution, 5, 39, 95, 125, signal-to-interference-and-noise ratio (SINR)
191, 212 model, 125
global optimal solution, 39 signal-to-noise ratio (SNR), 40
local optimal solution, 39 simulated annealing (SA), 7, 260, 263
optimization space, 100, 132, 252 single-input single-output (SISO), 40
core optimization space, 5, 125, 134, 146 smallest enclosing disk (SED), 193, 197, 212
orthogonal channel, 41, 65, 263 spatial reuse, 41
spectrum allocation, 244
parametric analysis (PA), 14, 22 spectrum sharing, 245
partitioning problem, 96 spectrum sharing problem, 245
partitioning variable, 4, 96, 119, 125 standard form, 12
path loss index, 155, 194, 207, 296 strong duality, 40
path-loss index, 103, 115, 233, 247 survival-of-the-fittest principle, 7, 261, 262,
per-node capacity, 285 269, 277
per-node throughput, 7, 285, 312, 313 synchronized power control, 284
330 Index
tabu search (TS), 7, 261, 263 unicast capacity, 284

three tangential supports, 133 unmanned autonomous vehicle (UAV), 211
throughput maximization problem, 5, 126, 146
time-dependent problem, 6, 212, 240 vertex-color problem, 289, 295
Tournament selection, 271
trace, 42 weak duality, 40
transmission range, 44, 103, 247, 285, 287 wireless ad hoc network, 7, 262, 272, 285, 313
transport capacity, 296 wireless charging vehicle, 152, 155
traveling salesman problem (TSP), 163 wireless energy transfer, 5, 151
magnetic resonant coupling, 151
ultra-wide band (UWB), 284 wireless power transfer, 155
unconstrained mobile base station (U-MB) wireless sensor network (WSN), 3, 5, 13, 151,
problem, 212, 223 192, 211

1107018803

Hochgeladen von

Dokumentinformationen

Originaltitel

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

1107018803

Hochgeladen von

Copyright:

Verfügbare Formate

9781107018808AR Engineering November 17, 2013 19:22 Char Count= 0

Applied Optimization Methods for Wireless Networks

There is a growing trend in applying optimization approaches to study wireless

Y. Thomas Hou is a Professor in the Bradley Department of Electrical and

University Printing House, Cambridge CB2 8BS, United Kingdom

Cambridge University Press is part of the University of Cambridge.

Part I Methods for Optimal Solutions 9

3 Convex programming and applications 38

3.8 Chapter summary 57

4 Design of polynomial-time exact algorithm 61

Part II Methods for Near-optimal and Approximation Solutions 93

6 Reformulation-linearization technique and applications 122

7 Linear approximation 148

7.3 Wireless energy transfer: a primer 153

8 Approximation algorithm and its applications - Part 1 191

9 Approximation algorithm and its applications Part 2 211

Part III Methods for Efficient Heuristic Solutions 243

11 Metaheuristic methods 262

Part IV Other Topics 281

For a graduate student (regardless of whether she has taken related OR or

We recognize that a single-volume book cannot possibly cover all tech-

Such a problem-oriented approach offers a convenient way of composing a

A highlight of this book is to include, in each chapter, a comprehensive presen-

xvii Copyright Permissions

Communications, vol. 8, no. 10, pp. 52905299, October 2009. Copyright

1.1 Book overview

The goal of this book is to offer a course to graduate students by demon-

Part I, consisting of three chapters, is devoted to optimization and designing

Structure of each chapter Each chapter starts with a brief overview of a

each chapter is not meant to be a replacement of formal courses in operations

To reflect state-of-the-art wireless network research rather than discuss some

Key characteristics of the book

Presents a collection of useful optimization techniques (one technique in

3 1.2 Book outline

1.2 Book outline

Chapter 2 reviews linear programming (LP) and shows how it can be

layering-based decomposition offers better insights and interpretations for

Part II of this book, consisting of Chapters 5 to 9, is devoted to techniques

Chapter 5 presents the branch-and-bound framework and shows how it can

5 1.2 Book outline

the worst-case complexity of a branch-and-bound-based method remains

multi-hop flow routing in the network are time-dependent. To address this

Chapter 10 presents an effective approach to address a class of mixed-

7 1.3 How to use this book

known metaheuristic methods (e.g., basic local search, simulated annealing

Chapter 12 presents an asymptotic capacity analysis for wireless ad hoc net-

1.3 How to use this book

This book is written as a textbook and is mainly aimed at graduate stu-

prepared a solutions manual for all end-of-chapter problems, which is available

I Methods for Optimal Solutions

2 Linear programming and

2.1 Review of key results in linear programming

Linear programming (LP) is a problem consisting of a linear objective function

12 Linear programming and applications

coefficient. The constant column vector with the ith component bi , 1 i m,

13 2.2 Case study

(MILP), which is not an LP problem. A special case of MILPs are integer

2.2 Case study: Lexicographic max-min rate allocation and

14 Linear programming and applications

15 2.3 System modeling and problem formulation

2.3 System modeling and problem formulation

We consider a two-tier architecture for wireless sensor networks. Figs. 2.1(a)

16 Linear programming and applications

k=i Vki + j =i Cij Vij + CiB ViB ei (i Sh )

Vij , ViB , l 0 (1 i, j N, j = i).